PHYSICAL REVIEW E 73, 011907 共2006兲

Response of electrically coupled spiking neurons: A cellular automaton approach

Lucas S. Furtado* and Mauro Copelli†

Laboratório de Física Teórica e Computacional, Departamento de Física, Universidade Federal de Pernambuco, 50670-901 Recife, PE,
Brazil
共Received 22 December 2004; revised manuscript received 6 September 2005; published 12 January 2006兲

Experimental data suggest that some classes of spiking neurons in the first layers of sensory systems are
electrically coupled via gap junctions or ephaptic interactions. When the electrical coupling is removed, the
response function 共firing rate vs. stimulus intensity兲 of the uncoupled neurons typically shows a decrease in
dynamic range and sensitivity. In order to assess the effect of electrical coupling in the sensory periphery, we
calculate the response to a Poisson stimulus of a chain of excitable neurons modeled by n-state Greenberg-
Hastings cellular automata in two approximation levels. The single-site mean field approximation is shown to
give poor results, failing to predict the absorbing state of the lattice, while the results for the pair approximation
are in good agreement with computer simulations in the whole stimulus range. In particular, the dynamic range
is substantially enlarged due to the propagation of excitable waves, which suggests a functional role for lateral
electrical coupling. For probabilistic spike propagation the Hill exponent of the response function is ␣ = 1,
while for deterministic spike propagation we obtain ␣ = 2 , which is close to the experimental values of the
1

psychophysical Stevens exponents for odor and light intensities. Our calculations are in qualitative agreement
with experimental response functions of ganglion cells in the mammalian retina.

DOI: 10.1103/PhysRevE.73.011907 PACS number共s兲: 87.19.La, 87.10.⫹e, 87.18.Sn, 05.45.⫺a

I. INTRODUCTION 共OSNs兲 to varying odorant concentration usually presents a

narrow dynamic range 关5,6兴. This is in contrast with
Unveiling how neuronal activity represents and processes the response observed in the next layers of the olfactory
sensory information remains a very difficult problem, despite bulb: both the glomerular 关7,8兴 and mitral cell 关9兴 responses
theoretical and experimental efforts undertaken by neurosci- present a broader dynamic range than the OSNs. In this
entists for the last several decades 共for a recent review, see case, the tightly packed unmyelinated axons of OSNs in the
关1兴兲. In this broad context, relatively little attention has been olfactory nerve are believed to interact electrically via ephap-
devoted to the question of how organisms cope with the sev- tic interactions 关10兴 共i.e., mediated by current flow through
eral orders of magnitude spanned by the intensities of sen- the extracellular space兲, as shown by Bokil et al. 关11兴. In
sory stimuli 关2兴. This astonishing ability is most easily re- particular, their results indicate that a spike in a single axon
vealed in humans by classical results in psychophysics 关3兴: can evoke spikes in all other axons of the bundle, suggesting
the perception of the intensity of a given stimulus is experi- that some computation is performed prior to the glomerular
mentally shown to depend on the stimulus intensity r as layer.
⬃log共r兲 共Weber-Fechner law兲 or ⬃r␣ 共Stevens law兲, where Motivated by these results, previous papers have shown
the Stevens exponent ␣ is typically ⬍1. Those laws have in through numerical simulations that electrical coupling
common the fact that they are response functions with broad among neurons indeed changes the response function in a
dynamic range, i.e., they map several decades of stimuli into way that is consistent with experimental results. Due to the
a single decade of response. coupling, stimuli generate excitable waves which propagate
through the neuron population. The interplay between wave
One would like to understand how this broad dynamic
creation and wave annihilation leads to a nonlinear amplifi-
range is physically achieved by neuron assemblies. Recent
cation of the spiking response, increasing the sensitivity at
experimental evidence suggests that electrical coupling low input levels and enhancing the dynamic range 关12,13兴.
among neurons in the early layers of sensory systems plays In one dimension, the robustness of the mechanism is at-
an essential role in weak stimulus detection. Deans et al. 关4兴 tested by the diversity of models employed: either the bio-
showed that electrical coupling is present in the mammalian physically realistic Hodgkin-Huxley equations 关12,14兴, a lat-
retina via gap junctions 共ionic channels that connect neigh- tice of nonlinear coupled maps 关13,15,16兴, or the Greenberg-
boring cells兲. In particular, the spiking response of ganglion Hastings cellular automata 共GHCA兲 关12,13兴 yield
cells to light stimulus changes dramatically when the gap qualitatively similar results. The same phenomenon has re-
junctions are genetically knocked out: both sensitivity and cently been observed in simulations with the two-
dynamic range are reduced 关4兴. dimensional GHCA 关17兴.
Another example comes from the olfactory system. The In this paper we calculate the response of excitable
spiking response of isolated olfactory sensory neurons GHCA model neurons 关18兴, where the bidirectional 共electri-
cal兲 coupling is modeled by a probability p of spike trans-
mission. While the uncoupled case p = 0 can be exactly
*Electronic address: [email protected] solved, the coupled case p ⬎ 0 is handled within two mean
†
Corresponding author. Electronic address: [email protected] field approximations, namely at the single-site and pair lev-

1539-3755/2006/73共1兲/011907共10兲/$23.00 011907-1 ©2006 The American Physical Society

L. S. FURTADO AND M. COPELLI PHYSICAL REVIEW E 73, 011907 共2006兲

els. The aim is to shed light on the analytical behavior of the undergoes an absolute refractory period before returning to
response function for the one-dimensional case, therefore rest.
building on previous efforts which have relied entirely on Whether the neurons are isolated or coupled is implicit in
numerical simulations. the definition of the supra-threshold stimulus. We assume
Our focus on the response of a continuously driven spa- external supra-threshold stimuli to be a Poisson process with
tially extended excitable system should be carefully con- rate r 共events per second兲. Hence at each time step an exter-
fronted with other recent studies, where the main interest has nal stimulus arrives with probability
been on phase transitions between an excitable and a self-
sustained collective state. For instance, the SIRS model of ␭共r兲 = 1 − e−r␶ 共1兲
epidemics in hypercubic lattices has been recently investi-
gated under the mean field and pair approximations 关19兴. In per neuron. Notice that ␶ = 1 ms corresponds to the approxi-
those contagion models, stationary self-sustained activity be- mate duration of a spike and is the time scale adopted for the
comes stable for sufficiently strong connection among neigh- time step of the model. The number of states n therefore
bors, a behavior which has been shown to be universal under controls the duration of the refractory period 共which corre-
very general assumptions 关20兴. Similar results have been ob- sponds to n − 2, in ms兲. In the biological context, r could be
tained for a variety of neuronal models, including collective related for example with the concentration of a given odorant
responses to a localized transient stimulus 关21,22兴, as well as presented to an olfactory epithelium 关5兴, or the light intensity
the emergence of sustained activity in complex networks stimulating a retina 关4兴. We shall refer to r as the stimulus
关22,23兴. rate or intensity.
While interesting in its own, the framework of stable- When electrically coupled, neurons at rest can also be
unstable collective transitions does not seem particularly stimulated by their neighbors. We define p and q as the prob-
suited for our modeling purposes. To account for sensory abilities that a resting neuron spikes as a consequence of
responses, the employed GHCA model is an excitable system transmission 共ionic current flow兲 from respectively one or
which always returns to its absorbing state in the absence of two spiking neighbors 关see Eq. 共3兲兴. We keep p and q as
stimulus; there are no phase transitions. The refractory pe- two independent parameters in most calculations to show
riod of the GHCA model neurons is absolute 共unlike, say, the robustness of some asymptotic results. In the simulations,
reaction-diffusion lattices兲, mimicking the deterministic be- we concentrate on the more physically intuitive choice of
havior of continuous-time systems like the Hodgkin-Huxley q = 1 − 共1 − p兲2, where the contributions from two spiking
equations or integrate-and-fire models 关14兴. The only source neighbors are independent.
of stochasticity of the model regards the firing of the neu- Let P共i兲
t 共k兲 be the probability that the ith neuron is in state
rons. Stimuli can come from spiking neighbors 共with prob- k at time t. Since the dynamics of the refractory state is
ability p兲 or from an “external” source, which is modeled deterministic, the equations for k 艌 2 are simply
by a Poisson process and represents sensory input. Therefore,
共i兲
in the limit p = 1 the dynamics is that of a deterministic ex- Pt+1 共2兲 = P共i兲
t 共1兲
citable lattice being stochastically stimulated, which casts
the problem into the framework of probabilistic cellular 共i兲
Pt+1 共3兲 = P共i兲
t 共2兲
automata 关24兴.
The paper is organized as follows. In Sec. II, the GHCA
rules are described; Sec. III contains the exact calculations ⯗
for the response of uncoupled neurons, while in Secs. IV and
V results for the coupled case are discussed in the mean field 共i兲
and pair approximations, respectively. Our concluding re- Pt+1 共n − 1兲 = P共i兲
t 共n − 2兲. 共2兲
marks are presented in Sec. VI.
To describe the coupling among first neighbors, let
P共i兲
t 共k , l , m兲 be the joint probability that sites i − 1, i and
II. THE MODEL i + 1 are respectively in the states k, l and m at time t. Fol-
lowing the definitions of ␭, p, and q above, the equation for
In the n-state GHCA model 关18兴 for excitable systems, P共i兲
t 共1兲 thus becomes
the instantaneous membrane potential of the ith cell
共i兲
共i = 1 , . . . , L兲 at discrete time t is represented by xi共t兲 Pt+1 共1兲 = 关1 − 共1 − ␭兲共1 − q兲兴P共i兲
t 共1,0,1兲 + 关1 − 共1 − ␭兲共1 − p兲兴
苸 兵0 , 1 , . . . , n − 1其, n 艌 3. The state xi共t兲 = 0 denotes a neuron
at its resting 共polarized兲 potential, xi共t兲 = 1 represents a spik-
ing 共depolarizing兲 neuron, and xi共t兲 = 2 , . . . , n − 1 account for
⫻ 冉兺
n−1

k⫽1
P共i兲
t 共1,0,k兲 +
n−1

兺 P共i兲t 共k,0,1兲
k⫽1
冊
the afterspike refractory period 共hyperpolarization兲. We em- n−1 n−1
ploy the simplest rules of the automaton: if xi共t兲 = 0, then + ␭兺 兺 P共i兲t 共k,0,l兲. 共3兲
xi共t + 1兲 = 1 only if there is a supra-threshold stimulus at site i; k⫽1 l⫽1
otherwise, xi共t + 1兲 = 0. If xi共t兲 艌 1, then xi共t + 1兲 = 关xi共t兲 + 1兴
mod n, regardless of the stimulus. In other words, the rules Finally, the dynamics for P共i兲
t 共0兲 can be obtained by the nor-
state that a neuron only spikes if stimulated, after which it malization condition

011907-2
RESPONSE OF ELECTRICALLY COUPLED SPIKING… PHYSICAL REVIEW E 73, 011907 共2006兲

FIG. 1. Response curves for 共a兲 n = 3 and 共b兲 n = 10 automata: FIG. 2. Response curves for 共a兲 n = 3 and 共b兲 n = 10 automata:
simulations 共symbols兲 and mean field approximation 关lines, accord- simulations 共symbols兲 and pair approximation 关lines, according to
ing to Eq. 共11兲兴. From bottom to top, p = 0, 0.3, 0.6, and 1, Eqs. 共17兲 and 共18兲兴. From bottom to top, p = 0, 0.3, 0.6, and
q = 1 − 共1 − p兲2. In the simulations, standard deviations over ten runs 1, q = 1 − 共1 − p兲2. The pair approximation eliminates the small-␭
are smaller than symbol sizes, so error bars are omitted in all fig- anomalies of the single-site solution, yielding excellent agreement
ures. Notice the negative slope and multi-valuedness of the single- with simulations for the extreme cases p = 0 and p = 1.
1
site approximation for p ⬎ 2 and ␭ 艋 0.
We are interested in obtaining the behavior of P共1兲 as a
n−1 function of ␭ 共or r兲. Note that P共1兲 coincides with the aver-
兺 P共i兲
t 共k兲 = 1, ∀ t,i, 共4兲 age firing rate per neuron 共measured in spikes per ms, ac-
k=0 cording to the choice of ␶兲 in the limit L , t → ⬁. In simula-
tions, firing rates have been calculated by division of the
which completes the set of equations for one-site total number of spikes in the chain by LT, where T
probabilities. ⬃ O共105兲 and L ⬃ O共105兲 were the typical number of time
It is reasonable to assume homogeneity in the system steps and model neurons employed. We define F共␭兲 ⬅ P共1兲
when L → ⬁, so we can drop the superscript 共i兲 in Eqs. as the response function of the system.
共2兲–共4兲 and in what follows. We also expect isotropy 共right- Due to the absolute nature of the refractory period,
left symmetry兲 in the probabilities: Pt共k , l兲 = Pt共l , k兲, the maximum firing rate of the model neurons is
Pt共k , l , m兲 = Pt共m , l , k兲, etc. Recalling the normalization con- Fmax ⬅ 1 / n, a result which is easily obtained ∀p , q by setting
dition 兺n−1 j1=0 Pt共j 1 , j 2 , . . . , j m兲 = Pt共j 2 , . . . , j m兲, one can rewrite ␭ = 1 in Eqs. 共5兲 and 共6兲. The dynamic range ␦␭ of the re-
Eq. 共3兲 as sponse curve F共␭兲 follows the definition commonly
employed in biology 关5,25兴:
Pt+1共1兲 = ␭Pt共0兲 + 2p共1 − ␭兲Pt共1,0兲
+ 共1 − ␭兲共q − 2p兲Pt共1,0,1兲. 共5兲 ␦␭ = 10 log10 冉 冊
␭0.9
␭0.1
, 共7兲

The stationary value of any joint probability will be de- where ␭x satisfies
noted by omitting the subscript t, thus P共•兲 ⬅ limt→⬁ Pt共•兲.
We start by solving Eqs. 共2兲 and 共4兲 in the stationary state, F共␭x兲 = xFmax . 共8兲
which together yield The dynamic range is therefore the number of decibels of
input which are mapped into the ⯝9.5 dB of output com-
P共0兲 = 1 − 共n − 1兲P共1兲, 共6兲 prised in the 关0.1Fmax , 0.9Fmax兴 interval 共see Fig. 3兲. In the
biological context of the model, it measures the ability of the
a result which is exact and holds ∀p , q. system to discriminate different orders of magnitude of

011907-3
L. S. FURTADO AND M. COPELLI PHYSICAL REVIEW E 73, 011907 共2006兲

FIG. 3. Linear-log plot of the response curve for n = 3 automata FIG. 4. Dynamic ranges 共triangles for ␦␭, circles for ␦r兲 as a
with p = q = 1 共filled circles兲, p = 0.5 and q = 1 − 共1 − p兲2 共open tri- function of the number of states of the GHCA, obtained from the
angles兲, and p = q = 0 共open circles兲. Lines correspond to the pair stationary solution of the pair approximation. Open 共filled兲 symbols
approximation. Horizontal lines are F = 0.1Fmax and F = 0.9Fmax, correspond to the p = q = 0 共p = q = 1兲 case. Inset: ␦␭ as a function of
vertical lines are ␭ = ␭0.1 and ␭ = ␭0.9, and arrows illustrate the dy- p for n = 10 for simulations 共dashed line兲 and pair approximation
namic range ␦␭ 关 Eq. 共7兲兴 for p = 0 and p = 1. The dynamic range of 共solid line兲. In spite of the underestimation of the response observed
a chain of neurons with deterministic spike propagation is about in Fig. 2, the pair approximation is able to reproduce the behavior
twice as large as that of its uncoupled counterpart. of the dynamic range as a function of the probability of spike
transmission.
stimulus intensity. We will show below that if one chooses to
calculate ␦r using rx ⬅ −␶−1 ln共1 − ␭x兲 instead of ␭x in Eq. 共7兲, and consists in discarding the influence of all neighbors in
results are essentially unchanged. the conditional probabilities 关26兴, thus Pt共j1 兩 j2 , . . . , jm兲
⬇ Pt共j1兲, which leads to
m
III. UNCOUPLED NEURONS
Pt共j1, . . . , jm兲 ⬇ 兿 Pt共jk兲. 共10兲
The uncoupled case p = q = 0 can be exactly solved by tak- k=1
ing the limit t → ⬁ in Eq. 共5兲 which, together with Eq. 共6兲,
In this approximation, Eq. 共5兲 becomes
yields
Pt+1共1兲 ⬇ Pt共0兲关␭ + 2p共1 − ␭兲Pt共1兲
␭
P共1兲 = f共␭兲 = . 共9兲 + 共q − 2p兲共1 − ␭兲Pt共1兲2兴, 共11兲
1 + 共n − 1兲␭
This linear saturating response is depicted for n = 3 共Figs. which, together with Eq. 共6兲, can be used to eliminate P共0兲
1–3兲 and n = 10 共Figs. 1 and 2兲, in complete agreement with and render P共1兲 = F共␭兲 implicitly through the relation
simulations. It belongs to the family of Hill functions defined
共1 − 2p兲F + 共2pn − q兲F2 + 共n − 1兲共q − 2p兲F3
by H␣共x兲 ⬅ Cx␣ / 共x0␣ + x␣兲, where the Hill exponent in this ␭⬇ . 共12兲
case is ␣ = 1. 关1 − 共n − 1兲F兴关1 − 2pF + 共2p − q兲F2兴
The dynamic range can be promptly calculated: As a consistency check, notice that setting p = q = 0 in Eq.
␦␭共n兲 = 10 log10关共1 + 9n兲 / 共1 + n / 9兲兴 and ␦r共n兲 = 10 log10兵ln共1 共12兲 recovers Eq. 共9兲 共in other words, mean field is exact for
+ 9 / n兲 / ln关1 + 1 / 共9n兲兴其, both of which rapidly converge to the uncoupled case, as it should兲. However, for 0 ⬍ p , q 艋 1,
10 log10共81兲 ⯝ 19 dB for moderate values of n 共see lower F共␭兲 as given by Eq. 共12兲 yields in general a poor agreement
curves in Fig. 4兲. As we shall see, the electrical coupling can with numerical simulations, as can be seen in Fig. 1 for dif-
lead to dynamic ranges typically twice as large. ferent values of p. When ␭ ⯝ 0, Eq. 共12兲 predicts F ⯝ ␭ / 共1
− 2p兲, which leads to obviously nonphysical results
IV. COUPLED NEURONS: MEAN FIELD for p 艌 21 共see leftmost part of Fig. 1兲. In particular, F共␭兲 is
APPROXIMATION multi-valued, leading to lim␭→0+ F ⫽ 0. The mean field result
therefore suggests a transition to an ordered state at ␭ = 0
As can be seen in Eq. 共5兲, Pt共1兲 depends on two- and which is simply forbidden by the automaton rules 关27兴. By
three-site probabilities, and in general k-site probabilities de- generalizing Eq. 共11兲, this failure to predict the absorbing
pend on up to 共k + 2兲-site probabilities. The dynamical de- state of the system can in fact be extended to regular lattices
scription of the system thus requires an infinite hierarchy of with coordination z, where the single-site approximation
equations. The mean field approximation at the single-site ␭→0
level corresponds to the simplest truncation of this hierarchy yields F ⯝ ␭ / 共1 − pz兲. Since this level of approximation is

011907-4
RESPONSE OF ELECTRICALLY COUPLED SPIKING… PHYSICAL REVIEW E 73, 011907 共2006兲

clearly not satisfactory for the calculation of the dynamic Since Pt共j , 0兲 depends on Pt共j − 1 , 0兲 and Pt共j − 1 , n − 1兲, and
range, a refinement is needed. Pt共0 , 0兲 depends, among others, on Pt共n − 1 , n − 1兲, all the
equations for two-site probabilities are in principle required
V. COUPLED NEURONS: PAIR APPROXIMATION
for the dynamical description of the system. Together
The pair approximation consists in keeping the influence with the equations for single-site probabilities, they form a
of only one neighbor in the conditional probabilities 关26兴, 共n2 + 3n兲 / 2-dimensional map whose stationary stable solu-
thus Pt共j1 兩 j2 , . . . , jm兲 ⬇ Pt共j1 兩 j2兲. In this case m-site prob- tion can be analytically studied. While the Appendix contains
abilities are reduced to combinations of up to two-site prob- details of the derivation of those equations, we discuss the
abilities. In particular, three- and four-site probabilities be- main results below.
come 关26兴 The main point to be noted is that the calculation of the
P共k,l兲P共l,m兲 stationary state presents additional difficulties when n 艌 4. In
P共k,l,m兲 ⬇ , 共13a兲 that case, the pair probabilities P共j , 0兲 with 2 艋 j 艋 n − 2 have
P共l兲
the same stationary value, but differ from P共n − 1 , 0兲. In par-
P共j,k兲P共k,l兲P共l,m兲 ticular, for p = q = 1 one obtains P共j , 0兲 = 0 关2 艋 j 艋 n − 2, see
P共j,k,l,m兲 ⬇ . 共13b兲 Eq. 共A10兲兴, which in turn leads to many other vanishing
P共k兲P共l兲
probabilities and gives the deterministic case a sparse sta-
It is therefore possible to rewrite Eq. 共5兲 in this approxima- tionary matrix 关see Eqs. 共A4兲, 共A5兲, and 共A8兲兴. Those terms
tion: do not exist for the n = 3 case, which makes its analysis con-
Pt+1共1兲 ⬇ ␭Pt共0兲 + 共1 − ␭兲Pt共1,0兲 siderably simpler. In either case, for n 艌 3 one obtains the

冉 冊
reasonable result P共n − 1 , 0兲 ⬇ P共1 , 0兲, the l.h.s. 共r.h.s.兲 being
Pt共1,0兲 associated to the end 共beginning兲 of an excitable wave front
⫻ 2p + 共q − 2p兲 . 共14兲
Pt共0兲 关see Eq. 共A12兲兴. Combining these results, a normalization
Equation 共14兲, on its turn, depends on Pt共1 , 0兲, whose evo- condition and the linearity of Eq. 共16兲 in Pt共0 , 0兲, we obtain
lution can be exactly obtained 共up to homogeneity and isot- 共see the Appendix兲
ropy assumptions兲:
Pt+1共1,0兲 = ␭Pt共n − 1,0兲 + p共1 − ␭兲Pt共n − 1,0,1兲
+ ␭共1 − ␭兲Pt共0,0兲 + p共1 − ␭兲共1 − 2␭兲Pt共1,0,0兲
− p 共1 − ␭兲 Pt共1,0,0,1兲.
2 2
共15兲
冋
P共0兲 − P共1,0兲 2 + 共n − 3兲 冉 共1 − p兲P共0兲 + 共p − q兲P共1,0兲
P共0兲 − pP共1,0兲
冊册
P共1,0兲P共0兲关P共0兲 − pP共1,0兲兴
With the help of the pair approximation in Eqs. 共13兲, Eq. 共15兲 ⬇ ,
becomes ␭P共0兲2 + p共1 − 2␭兲P共0兲P共1,0兲 − p2共1 − ␭兲P共1,0兲2

冉
Pt+1共1,0兲 ⬇ Pt共n − 1,0兲 ␭ + p共1 − ␭兲
Pt共1,0兲
Pt共0兲
冊 共17兲

冉
+ 共1 − ␭兲Pt共0,0兲 ␭ + p共1 − 2␭兲
Pt共1,0兲
Pt共0兲 which is valid ∀n 艌 3. Consider now the stationary state of
Eqs. 共6兲 and 共14兲. They can be combined in a quadratic equa-
− p2共1 − ␭兲
Pt共1,0兲2
Pt共0兲2
.冊 共16兲
tion for P共1 , 0兲, yielding

共2p − q兲P共1,0兲 ⬇ G±关P共0兲兴 ⬅ pP共0兲 ± 冑 P共0兲兵P共0兲关共n − 1兲p2 + 2p − q + ␭共n − 1兲共2p − p2 − q兲兴 + 共q − 2p兲其
共n − 1兲共1 − ␭兲
. 共18兲

Since P共1 , 0兲 must vanish ∀p , q in the limit ␭ → 0, G− is the until it converges to its stationary state. Despite the
only acceptable solution. growing number of equations with n, this method has the
The solution of Eqs. 共17兲 and 共18兲 determines P共0兲 advantage of avoiding unstable fixed points 关26兴
as a function of ␭. Instead of numerically solving 关Eqs. 共17兲 and 共18兲 can have more than one solution兴.
them, we iterate the 共n2 + 3n兲 / 2-dimensional map involving Once P共0兲 is known, the response P共1兲 = F共␭兲 is obtained via
the one- and two-site probabilities for each value of ␭ Eq. 共6兲.

011907-5
L. S. FURTADO AND M. COPELLI PHYSICAL REVIEW E 73, 011907 共2006兲

A. Deterministic spike propagation „p = 1…

Ordinary differential equations 共ODEs兲 are the standard

modeling tool in computational neuroscience. This is due to
the fact that, despite the stochastic nature of the opening and
closing of individual ionic channels, a neuron containing a
large number of such channels can very often be extremely
well described by a deterministic dynamics 关14兴 共an ap-
proach which has been established since the seminal work of
Hodgkin and Huxley 关28兴兲. In the present context, it is there-
fore important to address the case p = 1. This limit is consis-
tent with a variety of scenarios in which, in addition to the
dynamics of individual neurons, spike transmission is also
well described by deterministic behavior. Specifically regard-
ing our present study, deterministic spike transmission due to
electrical coupling has previously been employed in the lit-
erature to model axo-axonal interactions both via ephaptic FIG. 5. Log-log plot of the response curve for p = q = 1. Pair
interactions 共e.g., in the olfactory nerve 关11兴兲 and gap junc- approximation 共solid lines兲 and simulations 共symbols兲 follow a
tions 共e.g., in the hippocampus 关27,29兴兲. This is in contrast power law 共␣ = 2 兲 for weak stimuli, while finite size effects lead to
1

with, say, dendro-dendritic gap junctions or chemical syn- a linear response F ⯝ L␭ 共dotted lines兲 for ␭ ⱗ ␭c共L兲.
apses 共in the latter case, synaptic transmission can sometimes
be as low as 10% due to the inherent stochasticity in the ␭→0
process of neurotransmitter release 关14,30兴兲, where the p = 1 F共␭兲 ⯝ 冑2␭, 共19兲
limit can hardly be expected to apply. As we shall see in the a result that holds ∀n , q, as should be expected. This power
following, in addition to its biological relevance, the re-
law suggests a Hill function with ␣ = 21 , which is an excellent
sponse function for p = 1 also has a different characteristic approximation for F共␭兲 in the whole ␭ interval when n is
exponent which will help us understand the limiting behavior large. This result explains the doubling of the dynamic range
for p ⱗ 1. as compared to the uncoupled case and is reminiscent of
Figure 2 shows the excellent agreement between the pair reaction-diffusion processes modeled by lattice gases
approximation and the simulations when p = q = 1. One ob- 关31–34兴 and partial differential equations 关35兴. Since the Hill
serves that the response is particularly enhanced in the low function can be regarded as a saturating Stevens law, it is
stimulus range. This feature is best seen in the logarithmic interesting to note that the experimental values of the
scale of Fig. 3: in comparison with the uncoupled case Stevens exponents for light and smell intensities are respec-
p = 0, the effect of the electrical interaction is to increase the tively ␣ ⯝ 0.5 and ␣ ⯝ 0.6 关3兴.
sensitivity of the response for more than a decade, leading to Let us now consider a chain with finite L and a very small
a dramatic rise of the dynamic range. value of ␭ such that a single external stimulus occurs in a
For each value of n, we can thus obtain the stationary given time interval. In this case, the deterministic nature of
response F共␭兲 and the dynamic ranges ␦␭ and ␦r in the pair the propagation would lead to L spikes in the chain, while a
approximation. Even though the response curve changes single spike would be observed if the neurons were un-
considerably for varying n 共since F is bounded by ␭→0
Fmax = 1 / n, see Fig. 2兲, the dynamic range levels off coupled. One would thus have F ⯝ Lf, and since f ⯝ ␭ 关from
smoothly, as can be seen in Fig. 4. For increasing n, the ␭→0
dynamic range of the p = q = 1 case approaches twice the Eq. 共9兲兴 we obtain F ⯝ L␭. This corresponds to a linear
value for the uncoupled case. The fact that this result holds regime where excitable waves do not interact. If one in-
for both ␦r and ␦␭ can be understood on the basis of the creases ␭, waves will start annihilating each other, leading to
low-stimulus amplification, which plays the central role in the power law response of Eq. 共19兲, as can be clearly seen in
the phenomenon: in this regime ␭ is approximately linear in Fig. 5. For a given system size L, there is therefore a cross-
r. Should one choose a different relationship ␭共r兲, ␦r would over value ␭c共L兲 ⯝ 2 / L2 from a linear to a nonlinear re-
obviously have different values, but the drastic enhancement sponse. In an infinite chain, there is no linear response since
in the response due to the electrical coupling would not be for any nonzero stimulus rate two excitable waves will in-
affected. evitably interact.
In order to understand the low-stimulus amplification in- To assess the finite size effects in the biological context of
duced by the coupling, we have analyzed Eqs. 共17兲 and 共18兲 the model, we notice that the dynamic range will be affected
when ␭ ⯝ 0. Inspection of Fig. 2 and previous numerical only if ␭c共L兲 ⲏ ␭0.1, that is, for L ⱗ 20n. For neurons with
simulations 关12兴 suggests that P共1兲 ⯝ C␭␣, with ␣ ⬍ 1. This refractory periods of the order of tens of ms, neuronal assem-
ansatz can be inserted into Eqs. 共6兲, 共17兲, and 共18兲 for general blies with L ⲏ 103–4 should therefore be well approximated
p and q, yielding ␣ = 21 and p = 1 as solutions. Deterministic by the limit L → ⬁, as can be seen in Fig. 6. It is important to
spike propagation therefore leads to a power law response emphasize, however, that even small chains dominated by

011907-6
RESPONSE OF ELECTRICALLY COUPLED SPIKING… PHYSICAL REVIEW E 73, 011907 共2006兲

FIG. 6. Dynamic range as a function of the system size L for

p = q = 1. Lines are just guides to the eye. FIG. 7. Log-log plot of the response curve: pair approximation
共solid lines兲 and simulations 共symbols兲 with q = 1 − 共1 − p兲2 and
n = 3. For p ⱗ 1, there is a crossover between ␣ = 1 and ␣ = 2 . The
1
finite size effects still possess dynamic ranges which are sig-
nificantly larger than those of the uncoupled case. For horizontal arrow shows 0.1Fmax.
␭0.1 ⱗ ␭c共L兲, the dynamic range increases approximately
logarithmically with the total number of connected neurons, the simulations displayed in Fig. 7. Interestingly, such a
a result which holds for regular lattices in any dimension change in exponent for p ⬍ 1 seems to be absent from
关17兴. reaction-diffusion models in lattice gases 关31–34兴 as well as
partial differential equations 关35兴.
B. Probabilistic spike propagation „p Å 1… Thanks to the growing coefficient in Eq. 共20兲, for p ⱗ 1
For p ⫽ 1, communication between spiking and resting the proximity to the transition that occurs at p = 1 produces a
neurons may eventually fail. This provides us with the sim- crossover in the response from a linear to a square root be-
plest test under which the robustness of the mechanism for havior, dismissing the suspicion that a larger exponent might
dynamic range enhancement can be checked. From the bio- severely deteriorate the enhancement of the dynamic range
logical point of view, this regime could be useful for model- 共see Fig. 7兲. In particular, notice that, for p ⱗ 1, ␣ = 21 is the
ing networks of neurons connected by chemical synapses, for dominant exponent at F = 0.1Fmax, which is used to calculate
instance. the dynamic range 共see horizontal arrow in Fig. 7兲. This ex-
We start the analysis of the p ⫽ 1 case by noticing in Figs. plains the smooth monotonic increase in ␦␭ with p, as shown
2 and 3 that the agreement between simulations and the pair in the inset of Fig. 4, even though the exponent changes
approximation is better than the mean field results 共espe- discontinuously at p = 1. On the one hand, we observe that
cially in the low-stimulus region兲, but certainly not as good deterministic spike propagation 共p = 1兲 is certainly not essen-
as in the extreme cases p = 0 and p = 1. This inevitably affects tial for the enhancement of the dynamic range, in the sense
the estimation of the dynamic range via the stationary state that any p ⬎ 0 yields a better response than uncoupled neu-
of the pair approximation 共see below兲, but nonetheless al- rons. On the other hand, it is interesting to point out that, as
lows us to understand qualitatively how the response p is varied from 0 to 1, the increase in dynamic range is
changes as p varies. particularly pronounced for p ⲏ 0.9. This is in agreement
As pointed out in the preceding section, the dynamic with the conjecture that the reliability of electrical coupling
range is enhanced for p = 1 primarily due to the low-stimulus among spiking neurons could indeed play a significant role
amplification associated to the propagation of excitable in early sensory processing.
waves. As opposed to the deterministic case, however, for
p ⫽ 1 a single excitable wave traveling in an infinite
chain initially at rest will eventually die out. We should VI. CONCLUDING REMARKS
therefore expect a qualitative change in the response function We have calculated the collective response to a Poisson
for ␭ ⯝ 0. This is indeed confirmed by reinserting the ansatz stimulus of a chain of electrically coupled excitable neurons
P共1兲 ⯝ C␭␣ in Eqs. 共6兲, 共17兲, and 共18兲 without the constraint modeled by n-state Greenberg-Hastings cellular automata.
␣ ⬍ 1. In this case, the linear behavior suggested by the plots The single-site mean field approximation has been shown to
in Fig. 2 is easily confirmed: give poor results, failing to predict the absorbing state of the
␭→0
F共␭兲 ⯝ 冉 冊
1+p
1−p
␭, 共20兲
lattice in the absence of stimulus for p 艌 21 . The pair approxi-
mation yields a response curve which agrees reasonably well
with simulations in the whole stimulus range. It is interesting
which is again valid ∀n , q. Therefore, the low-stimulus re- to remark that the agreement is particularly good when
sponse for p ⬍ 1 is governed by ␣ = 1, which is confirmed by p = q = 1, a deterministic regime in which the GHCA lattice

011907-7
L. S. FURTADO AND M. COPELLI PHYSICAL REVIEW E 73, 011907 共2006兲

mimics a system of coupled nonlinear ODEs. This reinforces

an interesting perspective in the context of computational
neuroscience: the possibility of applying techniques from
nonequilibrium statistical mechanics to the study of spatially
extended nonlinear systems.
The enhancement of the dynamic range in the presence
of electrical coupling is due to low-stimulus amplification.
For uncoupled neurons 共p = 0兲 the response is governed by
the Hill exponent ␣ = 1, leading to a dynamic range of
⬃19 dB. For coupled neurons this value can be doubled in
the limit p = q → 1, when the Hill exponent becomes ␣ = 21 .
This value is close to Stevens exponents observed in psycho-
physical experiments of smell and light intensities. For 0
⬍ p ⬍ 1, the exponent remains ␣ = 1, but the dynamic range
increases smoothly, which can be understood on the basis of
the crossover behavior observed in the response function for FIG. 8. Experimental response curves 共normalized firing rate vs.
p ⱗ 1. light intensity兲 of retinal on-center ganglion cells in linear-log 共main
In the context of experiments at the cellular level, the plot兲 and log-log 共inset兲 scales 共data extracted from Fig. 6 of Ref.
enhancement of the dynamic range associated to an increase 关4兴兲. Filled 共open兲 circles are for WT 共Cx36-KO兲 mice, solid
in sensitivity is also observed in both the olfactory 关8兴 and 共dashed兲 lines show the results of the pair approximation, thus
L → ⬁, with p = q = 1 共p = q = 0兲. Upper curves are for n = 10, lower
visual 关4兴 systems. While the dynamic range of OSNs 共the
curves are for n = 3. The dot-dashed line corresponds to simulations
neurons which perform the initial transduction兲 is about with n = 10, p = q = 1, and L = 20.
⬃10 dB 关5,6兴, the glomeruli 共the next processing layer兲 have
dynamic ranges at least twice as large 关8兴. It remains to be
investigated experimentally whether this enhancement is in-
deed due to ephaptic interactions among the unmyelinated cells involves a complex intermediate neuronal circuit 共with
OSN axons in the olfactory nerve. bipolar, horizontal and amacrine cells 关36兴兲; and individual
Stronger experimental support for our conjecture on the neurons themselves can have subtle dynamical properties
role of electrical interactions is available for the mammalian 共such as adaptation, for instance兲. All these properties are
retina. Deans et al. 关4兴 have measured the firing rates of clearly absent from our simple one-dimensional CA model.
on-center ganglion cells for varying light intensity 共measured Yet it correctly predicts the reduction in the dynamic range
in isomerized molecules of rhodopsin per rod per second, of a neuronal system which loses electrical coupling among
or Rh* / rod/ s兲. The response curves have been obtained for its cells.
both wild type 共WT兲 mice as well as mice in which the In order to have a quantitative agreement between experi-
mental and theoretical curves, additional modeling efforts are
expression of the protein connexin36 共responsible for the
needed which incorporate specific details of the system un-
gap junction intercellular channels兲 has been genetically
der consideration. However, the response of simple models
knocked out 共Cx36-KO兲. The difference in the response
of excitable media remains an important subject to be stud-
curves can be seen in Fig. 8. They present the same qualita- ied, precisely because they have the potential to reveal
tive behavior of the curves shown in Fig. 3, exhibiting an simple mechanisms and scaling relations 关35兴 whose robust-
increase in dynamic range in the presence of electrical cou- ness can thereafter be subjected to further testing in experi-
pling: 14 dB for Cx36-KO and 23 dB for WT, values which ments and more detailed models. In this context, the simple
are of the same order as those of Fig. 6. In particular, the Greenberg-Hastings CA strikes an interesting balance, on the
exponent of the “coupled” 共WT兲 case is ␣ ⯝ 0.58 共see inset兲, one hand capturing essential features of collective neuronal
which is slightly larger than what is obtained in the pair dynamics, while on the other hand lending itself to analytical
approximation. techniques borrowed from nonequilibrium statistical me-
The quantitative agreement between the analytical and ex- chanics.
perimental curves is limited. On the one hand, the theoretical
n = 3 curve can provide a good fit of the Cx36-KO data for
p = q = 0, while the coupled case p = q = 1 does not adjust well
ACKNOWLEDGMENTS
to the WT data. For n = 10 and p = q = 1, on the other hand,
the WT data are well matched by simulations with a finite
L = 20 system 共staying below the L → ⬁ pair approximation兲, The authors would like to thank O. Kinouchi, S. G.
but for p = q = 0 the same n = 10 automata are unable to give a Coutinho, A. C. Roque, R. F. Oliveira, R. Publio, M. J. de
good fit of the Cx36-KO data. The difficulties of a quantita- Oliveira, and an anonymous referee for useful discussions
tive match are not surprising: the retina is organized in and comments. LSF is supported by UFPE/CNPq/PIBIC.
layers which have, to first order, a two-dimensional structure; MC acknowledges support from Projeto Enxoval 共UFPE兲,
signal processing from the photoreceptors to the ganglion FACEPE, CNPq, and special program PRONEX.

011907-8
RESPONSE OF ELECTRICALLY COUPLED SPIKING… PHYSICAL REVIEW E 73, 011907 共2006兲

APPENDIX: THE EQUATIONS FOR TWO-SITE Pt+1共1,1兲 = ␭2 Pt共0,0兲 + 2p␭共1 − ␭兲Pt共1,0,0兲

PROBABILITIES
+ p2共1 − ␭兲2 Pt共1,0,0,1兲

冉
1. Dynamics
Pt共1,0兲
In all derivations below, homogeneity and isotropy are ⬇ Pt共0,0兲 ␭2 + 2p␭共1 − ␭兲
assumed. The sign “⬇” denotes that the equality holds in the Pt共0兲
pair approximation 关Eqs. 共13兲兴. We start by writing down the
equation for Pt共0 , 0兲, which holds ∀n 艌 3: + p2共1 − ␭兲2
Pt共1,0兲2
Pt共0兲2
.冊 共A3兲
Pt+1共0,0兲 = Pt共n − 1,n − 1兲 + 2共1 − ␭兲关Pt共n − 1,0兲
Off-diagonal terms, on the other hand, ultimately depend on
− pPt共1,0,n − 1兲兴 + 共1 − ␭兲2关Pt共0,0兲
Pt共j , 1兲. For j = 2, the equation is simply
− 2pPt共1,0,0兲 + p2 Pt共1,0,0,1兲兴
Pt+1共2,1兲 = 共␭ + p − p␭兲Pt共1,0兲 + 共1 − ␭兲共q − p兲Pt共1,0,1兲
⬇ Pt共n − 1,n − 1兲 + 2共1 − ␭兲Pt共n − 1,0兲

冉
⫻ 1−p
Pt共1,0兲
Pt共0兲
冊
+ 共1 − ␭兲2 Pt共0,0兲 冋 冉
⬇ Pt共1,0兲 ␭ + 共1 − ␭兲 p + 共q − p兲
Pt共1,0兲
Pt共0兲
冊册 ,

冉 冊
共A4兲
Pt共1,0兲 Pt共1,0兲2
⫻ 1 − 2p + p2 . 共A1兲
Pt共0兲 Pt共0兲2 while for j 艌 3 one has
The dynamics for two-site probabilities in the refractory pe- Pt+1共j,1兲 = ␭Pt共j − 1,0兲 + p共1 − ␭兲Pt共j − 1,0,1兲

冉 冊
riod obey a simple recursive rule due to the deterministic
evolution of the automata: Pt共1,0兲
⬇ Pt共j − 1,0兲 ␭ + p共1 − ␭兲 . 共A5兲
Pt+1共j,k兲 = Pt共j − 1,k − 1兲, 2 艋 j,k 艋 n − 1. 共A2兲 Pt共0兲

On the one hand, diagonal terms Pt共j , j兲 with j 艌 2 recur- Finally, one needs equations for Pt共j , 0兲, j 艌 2 关recall Eq.
sively depend on Pt共1 , 1兲, whose dynamics can be written as 共16兲 for Pt共1 , 0兲兴. Like in Eq. 共A4兲, the case j = 2 must be
follows: considered separately:

Pt+1共2,0兲 = Pt共1,n − 1兲 + 共1 − ␭兲共1 − p兲Pt共1,0兲 + 共1 − ␭兲共p − q兲Pt共1,0,1兲

冉
⬇ Pt共1,n − 1兲 + 共1 − ␭兲Pt共1,0兲 共1 − p兲 + 共p − q兲
Pt共1,0兲
Pt共0兲
. 冊 共A6兲

For j 艌 3, on the other hand, one immediately obtains 2. Stationary state

We start by handling the case n ⬎ 4. In the stationary state,
Pt+1共j,0兲 = Pt共j − 1,n − 1兲 + 共1 − ␭兲关Pt共j − 1,0兲 the first term on the r.h.s. of Eq. 共A7兲 becomes, via recursive
iterations of Eq. 共A2兲,
− pPt共j − 1,0,1兲兴
P共j − 1,n − 1兲 = P共1,1 + n − j兲, ∀ j 艌 3. 共A8兲
⬇ Pt共j − 1,n − 1兲

冉 冊
The above result can on its turn be further developed by
Pt共1,0兲 means of Eq. 共A5兲 as long as 1 + n − j 艌 3, rendering the sta-
+ 共1 − ␭兲Pt共j − 1,0兲 1 − p , 共A7兲
Pt共0兲 tionary state of Eq. 共A7兲:

which completes the set of all pair equations. Upon 冉

P共j,0兲 ⬇ P共n − j,0兲 ␭ + p共1 − ␭兲
P共1,0兲
P共0兲
冊
iteration of Eqs. 共2兲, 共4兲, 共14兲, 共16兲, 共A1兲, and 共A7兲,
normalization conditions properly imposed in the initial + 共1 − ␭兲P共j − 1,0兲
conditions are naturally preserved. To determine the
response function P共1兲 = F共␭兲, we wait until the
共n2 + 3n兲 / 2-dimensional map reaches a stationary state
冉
⫻ 1−p
P共1,0兲
P共0兲
, 冊 3 艋 j 艋 n − 2. 共A9兲

for each value of ␭. We describe below how the analysis Notice that we have a nonhomogeneous set of n − 4
of the stationary state can be reduced to just two equations linear equations for x j ⬅ P共j , 0兲 : x j ⬇ axn−j + 共1 − a兲x j−1, where
关Eqs. 共17兲 and 共18兲兴. a ⬅ ␭ + p共1 − ␭兲P共1 , 0兲 / P共0兲 and x2 = P共2 , 0兲 accounts for the

011907-9
L. S. FURTADO AND M. COPELLI PHYSICAL REVIEW E 73, 011907 共2006兲

nonhomogeneity in the equations for x3 and xn−2. The solu- P共n − 1,0兲 ⬇ P共1,0兲, 共A12兲
tion of these equations is simply xn−2 ⬇ xn−1 ⬇ ¯ ⬇ x3 ⬇ x2,
as can be checked by inspection. The combination of Eqs. which completes the proof for n ⬎ 4. For n = 4, it suffices to
共A6兲 and 共A5兲 in the stationary state, on the other hand, leads invoke Eqs. 共A6兲 and 共A5兲 to show that P共2 , 0兲
to ⬇ J关P共1 , 0兲 , P共0兲兴. With this result, Eq. 共A12兲 holds for
n 艌 4. Finally, for n = 3, Eqs. 共A6兲 and 共A4兲 together also lead
P共j,0兲 ⬇ J关P共1,0兲, P共0兲兴
to Eq. 共A12兲.
⬅ P共1,0兲 冉 共1 − p兲P共0兲 + 共p − q兲P共1,0兲
P共0兲 − pP共1,0兲
冊,
Invoking the normalization condition Pt共0兲 = 兺n−1
one can deduce that, on the one hand,
j=0 Pt共j , 0兲,

2 艋 j 艋 n − 2. 共A10兲 P共0,0兲 = P共0兲 − 2P共1,0兲 − 共n − 3兲J关P共1,0兲, P共0兲兴.

共A13兲
One therefore obtains
P共n − 2,0兲 ⬇ P共n − 3,0兲 ⬇ ¯ ⬇ P共2,0兲 ⬇ J关P共1,0兲, P共0兲兴. On the other hand, in the stationary state Eq. 共16兲 depends
linearly on P共0 , 0兲, so it can be inverted, yielding 关after sub-
共A11兲
stitution of Eq. 共A12兲兴 P共0 , 0兲 as a function of P共1 , 0兲 and
Finally, notice that P共n − 1 , 0兲 can be obtained by combina- P共0兲. Equaling this function to Eq. 共A13兲, P共0 , 0兲 is elimi-
tion of Eqs. 共A7兲, 共A11兲, and 共A4兲: nated and one obtains Eq. 共17兲.

关1兴 G. Kreiman, Phys. Life Rev. 1, 71 共2004兲. 关20兴 P. S. Dodds and D. J. Watts, Phys. Rev. Lett. 92, 218701
关2兴 T. A. Cleland and C. Linster, Neural Comput. 11, 1673 共1999兲. 共2004兲.
关3兴 S. S. Stevens, Psychophysics: Introduction to its Perceptual, 关21兴 H. Hasegawa, Phys. Rev. E 67, 041903 共2003兲.
Neural and Social Prospects 共Wiley, New York, 1975兲. 关22兴 A. Roxin, H. Riecke, and S. A. Solla, Phys. Rev. Lett. 92,
关4兴 M. R. Deans, B. Volgyi, D. A. Goodenough, S. A. Bloomfield, 198101 共2004兲.
and D. L. Paul, Neuron 36, 703 共2002兲. 关23兴 T. I. Netoff, R. Clewley, S. Arno, T. Keck, and J. A. White, J.
关5兴 J.-P. Rospars, P. Lánský, P. Duchamp-Viret, and A. Duchamp, Neurosci. 24, 8075 共2004兲.
BioSystems 58, 133 共2000兲. 关24兴 G. Ódor, Rev. Mod. Phys. 76, 663 共2004兲.
关6兴 J.-P. Rospars, P. Lánský, P. Duchamp-Viret, and A. Duchamp, 关25兴 S. Firestein, C. Picco, and A. Menini, J. Physiol. 共London兲
Eur. J. Neurosci. 18, 1135 共2003兲. 468, 1 共1993兲.
关7兴 R. W. Friedrich and S. I. Korsching, Neuron 18, 737 共1997兲. 关26兴 A. P. F. Atman, R. Dickman, and J. G. Moreira, Phys. Rev. E
关8兴 M. Wachowiak and L. B. Cohen, Neuron 32, 723 共2001兲. 67, 016107 共2003兲.
关9兴 P. Duchamp-Viret, A. Duchamp, and M. Vigouroux, Chem. 关27兴 T. J. Lewis and J. Rinzel, Network Comput. Neural Syst. 11,
Senses 15, 349 共1990兲. 299 共2000兲.
关10兴 G. Lowe, Curr. Opin. Neurobiol. 13, 476 共2003兲. 关28兴 A. L. Hodgkin and A. F. Huxley, J. Neurophysiol. 117, 500
关11兴 H. Bokil, N. Laaris, K. Blinder, M. Ennis, and A. Keller, J. 共1952兲.
Neurosci. 21, RC173 共2001兲. 关29兴 R. D. Traub, D. Schmitz, J. G. R. Jefferys, and A. Draguhn,
关12兴 M. Copelli, A. C. Roque, R. F. Oliveira, and O. Kinouchi, Neuroscience 92, 407 共1999兲.
Phys. Rev. E 65, 060901共R兲 共2002兲. 关30兴 E. R. Kandel, J. H. Schwartz, and T. M. Jessell, eds., Essen-
关13兴 M. Copelli, R. F. Oliveira, A. C. Roque, and O. Kinouchi, tials of Neural Science and Behavior 共Appleton & Lange, Nor-
Neurocomputing 65–66, 691 共2005兲. walk, Connecticut 1995兲.
关14兴 C. Koch, Biophysics of Computation 共Oxford University Press, 关31兴 R. B. Stinchcombe, M. D. Grynberg, and M. Barma, Phys.
New York, 1999兲. Rev. E 47, 4018 共1993兲.
关15兴 S. M. Kuva, G. F. Lima, O. Kinouchi, M. H. R. Tragtenberg, 关32兴 M. D. Grynberg, T. J. Newman, and R. B. Stinchcombe, Phys.
and A. C. Roque, Neurocomputing 38–40, 255 共2001兲. Rev. E 50, 957 共1994兲.
关16兴 M. Copelli, M. H. R. Tragtenberg, and O. Kinouchi, Physica A 关33兴 M. J. de Oliveira, Phys. Rev. E 60, 2563 共1999兲.
342, 263 共2004兲. 关34兴 J. R. G. de Mendonça and M. J. de Oliveira, J. Stat. Phys. 92,
关17兴 M. Copelli and O. Kinouchi, Physica A 349, 431 共2005兲. 651 共1998兲.
关18兴 J. M. Greenberg and S. P. Hastings, SIAM J. Appl. Math. 34, 关35兴 T. Ohta and T. Yoshimura, Physica D 205, 189 共2005兲.
515 共1978兲. 关36兴 G. M. Shepherd, ed., The Synaptic Organization of the Brain
关19兴 J. Joo and J. L. Lebowitz, Phys. Rev. E 70, 036114 共2004兲. 共Oxford University Press, Oxford, 1998兲.

011907-10