Journal of Neuroscience Methods 220 (2013) 116–130

Contents lists available at ScienceDirect

Journal of Neuroscience Methods

journal homepage: www.elsevier.com/locate/jneumeth

Computational Neuroscience

A brief history of excitable map-based neurons and neural networks

M. Girardi-Schappo a , M.H.R. Tragtenberg a,∗ , O. Kinouchi b,c
a
Departamento de Física, Universidade Federal de Santa Catarina, 88040-900 Florianópolis, Santa Catarina, Brazil
b
Departamento de Física, FFCLRP, Universidade de São Paulo, 14040-900 Ribeirão Preto, São Paulo, Brazil
c
Center for Natural and Artificial Information Processing Systems, Universidade de São Paulo, Brazil

a r t i c l e i n f o a b s t r a c t

Article history: This review gives a short historical account of the excitable maps approach for modeling neurons and
Received 26 February 2013 neuronal networks. Some early models, due to Pasemann (1993), Chialvo (1995) and Kinouchi and
Received in revised form 19 July 2013 Tragtenberg (1996), are compared with more recent proposals by Rulkov (2002) and Izhikevich (2003).
Accepted 22 July 2013
We also review map-based schemes for electrical and chemical synapses and some recent findings
as critical avalanches in map-based neural networks. We conclude with suggestions for further work
Keywords:
in this area like more efficient maps, compartmental modeling and close dynamical comparison with
Difference equations
conductance-based models.
Neuron models
Coupled map lattices © 2013 Elsevier B.V. All rights reserved.
Neural networks
Excitable dynamics
Excitable media
Bursting
Map-based neuron
Map-based synapses

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
2. History of map-based neurons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
2.1. Early history . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
2.2. KT and KTz maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
2.3. Recent proposals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
2.3.1. The Rulkov model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
2.3.2. The Izhikevich model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
3. Summary of the addressed models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
4. Coupling maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
4.1. Types of coupling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
4.1.1. Nonbiologically motivated coupling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
4.1.2. Diffusive coupling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.1.3. Chemical synapses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.2. Applications of the synapse models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
4.2.1. Nonbiologically motivated . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
4.2.2. Diffusive coupling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
4.2.3. Impulsive coupling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
5. Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

∗ Corresponding author. Tel.: +55(48)3721 2854; fax: +55(48)3721 9946.

E-mail address: [email protected] (M.H.R. Tragtenberg).

0165-0270/$ – see front matter © 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.jneumeth.2013.07.014
 M. Girardi-Schappo et al. / Journal of Neuroscience Methods 220 (2013) 116–130 117

1. Introduction In order to deal with these drawbacks, we may opt to repro-

duce the dynamical behavior of neurons instead of reproducing the
The number of neurons in the human brain (86 billions (Azevedo involved biophysical mechanisms (mechanistic modeling). Starting
et al., 2009)) is nearly six times the number of trees in Amazonia. from a complicated HH-model, perhaps even a multicompartmen-
So, brain modelers must not forget that their job is comparable to tal model, we can perform a sequence of simplifications more or less
modeling patches of Amazonia, a staggering task. Since well devel- justified in order to obtain simpler models with fewer equations
oped models for single neurons already exist (Bower and Beeman, and lumped parameters (de Schutter, 2010). Examples of these
2003; Carnevale and Hines, 2006), with complex dendritic geome- reduced ODE based models are the FitzHugh–Nagumo excitable
try and tens of equations and parameters (Dayan and Abbott, 2001), neuron (FitzHugh, 1955; Nagumo et al., 1962), the Hindmarsh–Rose
it is not obvious what modeling level we should use in general. bursting neuron (Hindmarsh and Rose, 1984) and the Izhikevitch
As the proverbial forest not seen because of the trees, the model (Izhikevich, 2003). If we numerically integrate these ODEs
detailed study of singular neurons is an interesting subject per with the Euler method with a large time step, we can arrive at maps
se, but it is perhaps not necessary to understand the macroscopic with similar dynamical properties as the original systems (Rulkov,
dynamics and function of neuronal networks. Indeed, neuronal 2002; Izhikevich, 2004).
networks present collective phenomena, like synchronization Phenomenological modeling can start the other way around.
(Ibarz et al., 2011), waves (Cross and Hohenberg, 1993; Gua et al., This occurs because the phenomena to be studied set the level
2013; Wua et al., 2013) and avalanches (Chialvo, 2010). The history of modeling. Continuing with our forest modeling analogy, if our
of neuronal networks modeling is marked by this trade-off between interest is to study a single tree (or neuron), a biophysical HH-like
analytical/computational tractability and biological realism. modeling is desirable. But if we want to understand, say, the propa-
Since the connection between neurons is only sensitive to the gation of a forest fire, the modeling of the tree biophysics is mostly
action potentials that arrive at (electrical or chemical) synapses, the immaterial, and trees could be represented by sites with two states
important thing is to model the dynamics of these action potentials (0 = normal, 1 = burnt) (Christensen et al., 1993). In the same vein,
(their frequency or inter-spike intervals, if they come in bursts or McCulloch and Pitts (1943) proposed a binary threshold neuron,
single events, etc.). The emphasis in modeling the transmembrane whose state is given by 0 = rest and 1 = firing. With this method,
voltage dynamical behavior is called a phenomenological approach one starts with discrete time systems and searches for increas-
(in contrast to a mechanistic or biophysical approach), leading to a ing complexity until a dynamical model that reproduces the full
class of neuron models where map-based neurons are a new and phenomenology of neuronal dynamics is achieved.
promising tool. This paper gives a brief account of the pioneering Both approaches tend to converge at a middle ground formal-
proposals of neuronal maps due to Chialvo (1995), Pasemann (1993, ism: dynamical systems with discrete space and discrete time, but
1997), Kinouchi and Tragtenberg (1996) and Kuva et al. (2001) with continuous state variables, that is, dynamical maps (Ibarz
and compare them with more recent proposals due to Rulkov et al., 2011). Neuronal networks composed by these maps will be
(2001, 2002) and Izhikevich (2003), Izhikevich and Hoppensteadt an instance of coupled maps lattices (CMLs) (Kaneko, 1993, 1994).
(2004). In this paper, we review the early proposals of map-based neu-
There are two main routes to achieve map-based neuron ron models and the coupling schemes used to create such coupled
models with realistic dynamical properties. The first one is to maps networks. We also suggest some unexplored research topics
start from Hodgkin–Huxley (HH) type models, composed by cou- that could be examined both with conductance-based neurons and
pled nonlinear ordinary differential equations (ODE) which are map-based neurons, in order to stimulate computational neurosci-
already a simplification (due to spatial discretization) of full partial entists to use both approaches in a synergetic way.
differential equations that describe the neuron membrane. Com- This review is intended to organize the map-based neuron
putational neuroscience models based on the HH formalism, called models in sequential time (Section 2), highlighting two families
conductance-based neurons, is a well developed subject (Dayan of map-based modeling: (I) back from McCulloch and Pitts (1943)
and Abbott, 2001), but suffers from some limitations (de Schutter, approach to the Kinouchi and Tragtenberg (1996) and its exten-
2010): sion (Kuva et al., 2001) and (II) back from Chialvo (1995) to Rulkov
(2002) and Izhikevich (2003). Then, we perform a short compu-
tational comparison of the main neuronal models (Section 3). The
main purpose of Section 4 is to neatly list the most relevant cou-
• The HH-type models consist in several nonlinear coupled ODEs: plings which may be used to link maps into networks, only pointing
the simulation of a single neuron is orders of magnitude more to the most prominent results obtained with them. We finally ter-
costly than simplified neuron models; minate the review with some important remarks in Section 5.
• The biophysical data for constraining the parameter values (like
capacitances, axial resistances, density of ion channels, etc.) is
scarce and often obtained from diverse preparations (different 2. History of map-based neurons
animals, in vitro experiments, etc.). Most of the parameter ranges
used in simulations are simply informed guesses.
• The remaining parameter space of these models is huge and suffer This section is devoted to draw a line which connects the early
from the so called curse of dimensionality (Bellman, 2003). It is modeling of neurons, as state machines, to the most recent and
very costly to trace full phase diagrams, since with P parameters, powerful models, which are dynamical systems on their own, pre-
for example, we can have P(P − 1)/2 parameter planes. The mini- senting their most relevant features and reviewing some models
mal model of Hodgkin–Huxley, with only two active ion channels, that are still not well known, although they have been used recently
has at least P = 40 parameters (Dayan and Abbott, 2001). to model neural networks.
• The set of parameters to be used for reproducing a given firing The generalized mathematical form of any map is:
pattern is subdetermined. This means that the same dynami- X(t  X(t)),
 + 1) = F(  (1)
cal behavior can be achieved by different sets of parameters.
Adjusting these parameters to the known neuron behavior is sus- where F : Rn → Rn is any vector function 
and we are assuming 

that the curve given by the set of values X(0), 
X(1), 
X(2), ...
ceptible of overfitting: the model reproduces the given data but
do not generalizes well, for example, for different input situations. defines the temporal evolution the system. In the case of neurons,
118 M. Girardi-Schappo et al. / Journal of Neuroscience Methods 220 (2013) 116–130

each component of the vector X(t)  accounts for a relevant neuronal xN (t) x2(t) x1(t)
quantity. ... i= j ... ... x(t − 1) x(t)
Generally, the first component, X1 (t), is the membrane poten- ···
... 1 ...
(a) GiN Gi2
tial (the fast variable) and the second component, X2 (t), may be the
slow inward and outward currents or an auxiliary variable. When
Gi1 (b)
present, the third component, X3 (t), accounts for the slow currents. W (1) W (0)
For the sake of simplicity, we define x(t) ≡ X1 (t), y(t) ≡ X2 (t) and − H ... ...
1
z(t) ≡ X3 (t) throughout this paper. The subscript index is reserved − H ... ...
1 1
to identify single elements of a network. For example, xi (t) is the xi (t + 1) x(t + 1)
membrane potential of the ith element of the network at time t.
Fig. 1. (a) Scheme of a single-layer perceptron (Eq. (2)) and (b) the so-called dynam-
2.1. Early history ical perceptron (Eq. (6) with  = 1). In both cases, the output may be calculated by a
continuous function or by a discrete step function.

McCulloch and Pitts (1943) formal neuron can be viewed, if cou-

pled to N presynaptic neurons with parallel update, as a discrete
time dynamical system (Little, 1974): not to continuous maps, that is, the state variables represents
⎛ ⎞ spikes, not continuous voltage potentials. Starting from the decade

N of 1980, modelers substituted the discontinuous step or sign func-
xi (t + 1) =  ⎝ Gij xj (t) + H + Ii (t)⎠ (2) tions by continuous ones:
j=
/ i


where the Heaviside (step) function gives (y) = 1 if y ≥ 0 (zero oth- x(t + 1) = F W (n) x(t − n) + H + I(t) , (6)
erwise), Ii (t) is the external input, −H is a firing threshold and the
n=0
Gij are the coupling parameters.
The isolated McCulloch–Pitts neuron has no intrinsic dynamics where, for example, the transfer function is a logistic function
(notice that the input sum is over j = / i). However we can introduce F(y) = [1 + exp(− y/T)]−1 (Aihara et al., 1990) or a hyperbolic tan-
a dependence on the past values of its variable, as in the Caianello gent F(y) = tanh(y/T) (Hopfield, 1984), where 1/T is a gain parameter.
(1961) equations: Notice that in the chaotic neuron of Aihara et al., x(t + 1) depends
⎛ ⎛ ⎞ ⎞ on all its previous values ( = t). Hence, for a given , x(t + 1) =

 
N
F (x(t), x(t − 1), . . ., x(t − )) is a  + 1 dimensional map.
xi (t + 1) =  ⎝ ⎝W (n) xi (t − n) + Gij xj (t)⎠ + H + Ii (t)⎠ ,
(n)
i Fig. 1 illustrates the differences between Eqs. (2) and (6) for  = 1.
n=0 j=
/ i
The case  = 0 with a logistic function F(y) and W(0) = 1, that
(3) is, x(t + 1) = F (x(t) + H) was examined by Pasemann (1993). In
(n) the plane (T, H) it presents only three behaviors: a fixed point
where now the Wi parameters weight the contributions of the
delayed xi (t − n) values within a memory window . So, the isolated phase, a period-two phase and a bistability (two fixed points) phase
neuron (Gij = 0) can present an interesting dynamical behavior. (Pasemann, 1993; Kinouchi and Tragtenberg, 1996). A chaotic ver-
Indeed, Nagumo and Sato (1972) studied an isolated formal neu- sion of this neuron map has also been studied by Pasemann (1997).
ron with an exponential decay of refractory influence, that is, Due to a lucky accident, the case  = 1 implements a powerful
excitable element with a very rich behavior. We call this sys-


tem a second order dynamical perceptron or the Kinouchi and
x(t + 1) =  W (n) x(t − n) + H + I(t) , (4)
Tragtenberg (1996) map (also known in the statistical mechanics
n=0 literature as the YOS map (Yokoi et al., 1985; Tragtenberg and Yokoi,
where W (n) = −˛w−n , with the scaling parameter ˛ > 0 and decay 1995)). This excitable neuron model is reviewed in the next section.
constant w > 1. It has been shown that almost all solutions of the So far all the models follow the same principle: they are built
Nagumo–Sato model are periodic and form a complete devil stair- directly from discrete time dynamics. However, neuronal mod-
case – chaotic solutions lie at the complementary zero measure els can be built the other way around: starting from continuous
cantor set (Aihara and Suzuki, 2010). time differential equations, like the HH model, and then simplify-
In the statistical physics community, due to symmetry motiva- ing them to keep only their wanted features. A result of this kind
tions, it is common to scale the neuron output as a sign function of simplification is proposed by Chialvo (1995) to study excitable
with values ±1, that is, S(y) = 2(y) − 1. For example, Horn (1993) systems (and, in particular, neural excitability):
studied a dynamical element which is a McCulloch–Pitts neuron
with a dynamical threshold H(t), given by x(t + 1) = [x(t)]2 exp[y(t) − x(t)] + I(t),
⎛ ⎞ (7)
y(t + 1) = ay(t) − bx(t) + c,

N
xi (t + 1) = S ⎝ Gij xj (t) − Hi (t) + Ii (t)⎠ ,
(5) where a, b and c are parameters and I(t) may account for a bias mem-
j=
/ i
brane potential or external input. Chialvo inspired more recent
Hi (t)
Hi (t + 1) = + bxi (t + 1), works, like the Rulkov (2002) and the Izhikevich (2003) models,
c
both studied in Section 2.3.
where b and c > 1 are parameters. Horn showed that, for large
b and c ≈ 1, the neuron presents an extended recovery period
which mimics regular spiking neurons. Bursting neurons can be 2.2. KT and KTz maps
simulated by substituting the last term in the H equation by
bxi (t + 1)xi (t)xi (t − 1) · · · xi (t + 1 − m) which ensures that the thresh- The case  = 1 with F(y) = tanh(y/T) was extensively studied in
old grows only after the neuron has fired m times. the context of Statistical Mechanics and resulted in many phase
Since the x(t) variable in all these models has a binary set of val- diagrams (Tragtenberg and Yokoi, 1995). In order to build on these
ues, this kind of formal neurons corresponds to cellular automata, results, Kinouchi and Tragtenberg (1996) imposed, in Eq. (6), the
 M. Girardi-Schappo et al. / Journal of Neuroscience Methods 220 (2013) 116–130 119

0.1 are present in Hodgkin–Huxley-like neurons and in experimental

Stability Limit setups (Kinouchi and Tragtenberg, 1996):

0.05
Bistable • Neuron A: Silent neuron – it will not fire any action potential,
regardless of the external stimulus intensity – no bifurcation
occurs;
• Neuron B (1 and 2): Pacemaker neuron – it is an autonomous
2 Fixed
0 Fast Spiking
H

oscillator, although external stimulation may bring the neu-

Points B1 ron to outside of the bulb in Fig. 2 temporarily via a
E B2 A Subcritical Neimark–Sacker bifurcation (B1) or a Supercritical
-0.05 D C1 Neimark–Sacker bifurcation (B2);
C2 • Neuron C (1 and 2): Excitable neuron – it will fire only for a stim-
1 Fixed Point ulus greater than some threshold; the stimulus takes the neuron
inside the bulb in Fig. 2 through a Subcritical Neimark–Sacker
-0.1
0.2 0.3 0.4 0.5 0.6 bifurcation (C1) or a Supercritical Neimark–Sacker bifurcation
T (C2). If a neuron C1 is adjusted over the Bistable region, then
external stimuli may switch between fast spiking and fixed point
Fig. 2. KT model phase diagram for K = 0.6. It allowed authors to list five different response (a phenomenon known as pacemaker activity annihila-
neuronal behaviors, labeled from A to E, discussed in the text. The bistable region
tion);
corresponds to coexistence of oscillatory with fixed point behaviors.
• Neuron D: presents a single fixed point; however, a strong enough
external current, I(t), may transform it into a bistable neuron with
change of parameters W(0) ≡ 1 and W(1) ≡ − K. Rewriting Eq. (6) two stable fixed points;
leads us to the KT model: • Neuron E: presents two stable fixed points even without exter-
nal input; the neuron may switch between states through some
x(t) − Ky(t) + H + I(t)
x(t + 1) = tanh , stimulation I(t).
T (8)
y(t + 1) = x(t), Type C1 neuron, specially, presents a rich repertoire of excitable
responses (see Fig. 3). It also provides insights about the math-
where −1 < x(t) < +1 represents the actual membrane potential of ematical nature of transient oscillations – not explained, but
the neuron at time t – measured in timesteps (ts). The term I(t) corre- experimentally observed by Morris and Lecar (1981): all behaviors
sponds to an external input. In Section 4, the coupling is done via I(t) of type C1 are achieved because the neuron lies near a bifurcation
as well. Both x(t), I(t) and t may be conveniently rescaled to any unit point.
system, like millivolts or milliseconds (ms). The authors assumed The KT model also presents a well defined excitability, as shown
1 ts = 0.1 ms whilst the membrane potential may be rescaled to fit in the last panel of Fig. 3 – in which the neuron receives a delta
a Hodgkin–Huxley model, for instance, by relating the value of the current input I(t) = I0 ıt,t0 in t0 = 0 with increasing intensity I0 . The
membrane potential at the fixed point and at the peak of the spike circle-dash line shows the approximate threshold Is : any stronger
in both models. stimulus causes a spike. The nullclines of the model are of the same
The authors reinterpreted the statistical mechanical phase dia- shape as those of the FitzHugh–Nagumo model (see Fig. 4).
grams, as shown in Fig. 2, proposing that the relevant neuronal Besides excitability, the KT model also presents regular and
behaviors are in the parameter subspace given by 0.5 ≤ K ≤ 1, with chaotic autonomous behavior (Fig. 5). The authors also proposed
T and H near the phase transitions from fixed point to oscillatory a mechanism to generate bursts based on the Hindmarsh–Rose
behavior (i.e. fast spiking). model (Kinouchi and Tragtenberg, 1996): by letting the parameter
The stability limit of the model, the curve Hs (K, T), separating H ≡ z(t), in Eq. (8), oscillate slowly in time, with
the oscillating and the fixed point phases, may be calculated via a
simple linear stability analysis, which yields: z(t + 1) = (1 − ı)z(t) − (x(t) − y(t))2 , (11)

where ı and  are parameters that control the inward and the
Hs± (K, T ) = T atanh(xs∗ ) − (1 − K)xs∗ , (9)
outward ionic currents, respectively. This modification allows the
where xs∗ is the fixed point over the line Hs (K, T), given by: neuron to go back and forth inside the bulb in Fig. 2. This mechanism
generates bursts when the transition induced by the z(t) dynamics
⎧  is via a Subcritical Hopf Bifurcation (i.e. through the bistable region
⎪
⎪ T
⎨ ± 1 − 1 − K , if 0 < K ≤ 0.5 in Fig. 2).
∗
xs =  . (10) Furthermore, Veiga and Tragtenberg (2001) showed that KT
⎪
⎪ T
⎩± 1− , if 0.5 < K ≤ 1
model, Eq. (8), may present stochastic resonance (which is the
K occurrence of spiking due to subthreshold stochastic stimula-
tion). Beyond the usual stochastic resonance, the authors achieved
Differently from the statistical mechanical model, the KT model autonomous and aperiodic stochastic resonance as well.
may be extended for T < 0 and K < 0. It would result in a slightly Kuva et al. (2001) modified the z(t) dynamics (from Eq. (11))
different xs∗ , although the analysis would keep the same (Kinouchi of the three variables KT model into a simpler version of the
and Tragtenberg, 1996). Thus, we will focus on positive parameters. neuron, which also presents bursts. We name it the KTz model.
Notice that the parameter H may be redefined as H̃ = H + I(t), The equations are:
so the effect of an external input is to drag the solution x(t) for
Eq. (8) along any vertical line (with fixed T) on Fig. 2. This is the x(t) − Ky(t) + z(t) + I(t)
x(t + 1) = tanh ,
basic mechanism of the KT model excitability, in which I(t) is such T
that H + I(t) > Hs (K, T) – assuming the neuron is adjusted below the (12)
y(t + 1) = x(t),
bulb of Fig. 2. This allowed the authors to list five distinct neu-  
ronal behaviors (labeled from A to E in Fig. 2). These behaviors z(t + 1) = 1 − ı z(t) −  (x(t) − xR ) ,
120 M. Girardi-Schappo et al. / Journal of Neuroscience Methods 220 (2013) 116–130

(a) fixed point (b) transient osc. (c) tonic spiking

I(t)
x(t)
20 ts

(d) nerve blocking (e) bistability (f) excitability

threshold

Fig. 3. Excitable response of KT type C1 model. The solution x(t) (◦) corresponds only to the circle points (the lines are just guides so one can follow the sequence in which
the map evolves through time). Solid line is external input and the bolder line segment corresponds to 20 ts. External current can be a delta pulse, I(t) = I0 ıt,t0 , a step current
with intensity I0 or a current ramp I(t) = a(t − t0 ). Parameters are K = 0.6, T = 0.35 and (a) H = −0.06, I0 = 0.04; (b) H = −0.06, I0 = 0.0475; (c) H = −0.06, I0 = 0.06; (d) H = −0.06,
a = 0.001; (e) H = −0.014, I0 = ±0.1; (f) H = −0.04 with, from the bottom to the top curve, 0.04 ≤ I0 ≤ 0.18 in steps of 0.01. The green curve stands on the excitability threshold.
(For interpretation of the references to colour in this figure legend, the reader is referred to the web version of the article.)

0.8
H=0 H = 0.03 H = 0.08
0.4 K = 0.6
0 T = 0.32
y

-0.4
x(t+1) = x(t)
-0.8 y(t+1) = y(t)
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
x
Fig. 4. Effect of varying parameter H on the nullclines of the KT model. Notice the similarity with the nullclines of the FitzHugh–Nagumo model.

Fig. 5. Left: Fast spiking behavior (inside the bulb in Fig. 2). Right: Chaotic spiking.

where x(t) is the membrane potential of the neuron, y(t) is a recov- We classified the KTz excitable behaviors according to
ery variable and z(t) is the slow total ionic current. The K and T Izhikevich and Hoppensteadt (2004), searching for qualitative sim-
parameters control the fast spiking dynamics, as explained above. ilarities between their model and the KTz model. As examples, we
The parameter ı is the inverse recovery time of z(t),  and xR con- present 15 out of the 20 excitable behaviors described by Izhikevich
trol the slow spiking and bursting dynamics (Kuva et al., 2001). In and Hoppensteadt (2004).
practice, though, the roles of ı,  and xR are illustrated in Fig. 6. Although it does not present every excitable behavior, it is a min-
KTz autonomous behaviors are listed in Fig. 7 whilst the imal model, based on only five parameters (K, T, ı,  and xR ). Besides,
excitable behaviors are in Fig. 8. A phase diagram for  = ı = 0.001 the KTz map (Eq. (12)) has no singularities and the functions are not
and K = 0.6 is given by Kuva et al. (2001) (see Fig. 9). The fixed piecewise defined (as the more recent proposals discussed in the
point stability of this model has been studied by Copelli et al. next section). Furthermore, the great symmetry of this model facil-
(2004). itates its phase plane analysis, which may be accomplished through
 M. Girardi-Schappo et al. / Journal of Neuroscience Methods 220 (2013) 116–130 121

1 (a) δ = 0.01 (b) δ = 0.1

0.5 I = 0.3δ t,0 I = 0.3(δ t,0+δ t,30)

x(t)
0 Fixed point
-0.5
-1
1 (c) δ = 0.01 (d) δ = 0.001
x(t)0.5 I = 0.3(δ t,0+δ t,30) I = 0.3(δ t,0+δ t,30)
0 K=0.6; T=0.35; λ=0.1; xR=-0.9
-0.5
-1
0 20 40 t 60 80 0 20 40 t 60 80

1 (e) λ= 0 (f) λ= 0.006

0.5 I = 0.3δ t,0 I = 0.3δ t,0
x(t)

0
-0.5
-1
1 (g) λ= 0.01 (h) λ= 0.1
0.5 I = 0.3δ t,0 I = 0.3δ t,0
x(t)

0 K=0.6; T=0.35; δ=0.1; xR=-0.9

-0.5
-1
0 20 40 t 60 80 0 20 40 t 60 80

1 (i) xR = -0.9 (j) xR = -0.6

0.5 K=0.6; T=0.35; δ=0.1; λ=0.1;
x(t)

0
-0.5
-1
1 (k) xR = -0.5 (l) xR = -0.2
0.5
x(t)

0
-0.5
-1
0 500 1000 t 1500 2000 0 500 1000 t 1500 2000

Fig. 6. (a) to (d): ı control the length of the refractory period; (e) to (h):  control the damping of oscillations; (i) to (l): xR control the bursting dynamics. Parameters’ values
are listed in the panels.

lots of detailed phase diagrams from which we can infer KTz behav- where ˛, , ˇ, ˇe ,  and  e are the six adjustable parameters. The
ior (Tragtenberg and Yokoi, 1995; Kinouchi and Tragtenberg, 1996; map consists of a fast dynamics in x(t) and a slow dynamics in y(t)
Kuva et al., 2001; Copelli et al., 2004). when  «1.
By the analysis of the nullclines of the fast and of the slow sub-
2.3. Recent proposals systems, one gets to the excitability threshold,  th , in the ˛ × 
plane, which corresponds to a Subcritical Hopf Bifurcation (Rulkov,
During the last decade, three other maps have been proposed: 2002):
the Rulkov model (Rulkov, 2002), the Izhikevich model (Izhikevich √
and Hoppensteadt, 2004) and the Courbage–Nekorkin–Vdovin th = 2 − ˛. (15)
model (Courbage et al., 2007). These and some other models have
Eq. (15) is plotted in Fig. 10 together with the limit between
been reviewed elsewhere (Courbage and Nekorkin, 2009; Ibarz
bursting and spiking phases (which was determined computation-
et al., 2011), so we are going to present only a brief description of the
ally by the author). There are also a chaotic (Rulkov, 2001) and
first two models because they are more easily found in literature.
a supercritical (Shilnikov and Rulkov, 2004) variant of the Rulkov
model. The excitability of the latter is due to a supercritical Hopf
2.3.1. The Rulkov model
bifurcation, hence its name.
The model proposed by Rulkov (2002) is, originally:
x(t + 1) = F(x(t), y(t) + ˇ + ˇe I(t)) 2.3.2. The Izhikevich model
, (13)
y(t + 1) = y(t) −  [x(t) + 1] +  [ +  e I(t)] Firstly proposed as an ordinary differential equation (ODE)
of continuous time (Izhikevich, 2003), the map-based Izhike-
where F(u, v) is a piecewise defined function given by: vich model is the following discretization (Izhikevich and
⎧ ˛ Hoppensteadt, 2004):
⎪
⎨ 1 − u + v ifu ≤ 0
F(u, v) = , (14) x(t + 1) = 0.04x2 (t) + 6x(t) + 140 − y(t) + H + I(t)
⎪ ˛+u if0 < u < ˛ + v , (16)
⎩ y(t + 1) = a[bx(t) − y(t)]
−1 ifu≥˛ + v
122 M. Girardi-Schappo et al. / Journal of Neuroscience Methods 220 (2013) 116–130

Table 1
= 100 ts (a) Comparison of CPU time, in nanoseconds [ns], of 1 ts for each map-based neuron
model. We also present the CPU time of a HH-like neuron modeled by Cymbalyuk
and Calabrese (2000) and integrated using a 4th order Runge-Kutta method. All
these models are adjusted in bursting activity. Here, T is the CPU time (real time),
(b)  is the conversion factor between ms and ts and TM is the model elapsed time.
Conversion factors are from Izhikevich and Hoppensteadt (2004), Ibarz et al. (2011),
Kuva et al. (2001). These factors generally take into consideration the empirical fact
(c) that 1 spike takes, approximately, 1 ms.

Model T 32bit [ns] T 64bit [ns]  [ms/ts] TM [ms]

x(t)

Izhikevich 9.6 ± 0.6 10.7 ± 0.7 1a 104

(d) Rulkov 11.6 ± 0.6 11.5 ± 0.5 0.5 5 × 103
KTz 30.9 ± 0.9 18.9 ± 0.7 0.1 103
HH 210 ± 4 255 ± 5 0.01 102
(e) a
Sometimes, the Izhikevich conversion factor is  = 0.5 ms/ts.

(f)
The Izhikevich model, with only 13 FLOP, may describe 20 different
behaviors (Izhikevich, 2004).
(g) Since the amount of FLOP is not a precise measure of the perfor-
mance of any equation (as it discard anything else than sums and
multiplications, like divisions, memory assignments, function calls,
Timesteps etc.), we decided to do a benchmark test with the last three pre-
Fig. 7. Examples of KTz model autonomous behaviors. When not specified, K = 0.6,
sented models by averaging the CPU time, T, in nanoseconds (ns),
T = 0.35 and ı =  = 0.001. (a) Subthreshold oscillations (xR = −0.5, T = 0.45); (b) each model takes to evaluate during Tts = 10000 ts. We repeat this
chaotic bursting (xR = −0.4, T = 0.322); (c) slow bursting (xR = −0.6); (d) fast burst- evaluation S = 10000 times to reduce the variation of the measures.
ing (xR = −0.45); (e) cardiac spikes (xR = −0.5, T = 0.25); (f) slow spiking (xR = −0.62, Thus, 1 ts takes, in average, T = T/(Tts S) real life nanoseconds to
ı =  = 0.003); (g) fast spiking (xR = −0.2, T = 0.45).
complete. The results are in Table 1 for codes running in 32-BIT and
64-BIT assemblies (both in a 64-BIT CPU). In this table, the elapsed
with the reset, if x(t) ≥ 30: time for each model, TM , is calculated by TM = Tts , where  is the
amount of milliseconds for each timestep of the considered model
x(t + 1) = c
, (17) [ms/ts].
y(t + 1) = y(t) + d Changing from 32 bits to 64 bits makes the KTz neuron
reduce to almost half of its CPU time, getting closer to the other
where a, b, c, d and H are parameters, I(t) is an input current
models’ performance, whilst the others keep with their same CPU
(whether synaptic or external) and t is the time in milliseconds,
time. This improvement is probably due to the numeric calculation
assumed. The parameters a and b control balance of the slow
of the hyberbolic tangent function, which may demand high com-
dynamics of the in y(t).
putational efforts when its argument goes to ±∞ (Cody and Waite,
Fig. 11 shows the bifurcation diagram in the plane c × H for
1980), although we do not know any work comparing the perfor-
a = 0.02, b = 0.25 and d = 0. The bifurcations undergone by the Izhike-
mance of such algorithms in different computer architectures.
vich model depend strongly on parameters.
Nevertheless, one must remind that the choice of the model
The author has proposed this model as a reduced
depends strongly on the problem one is trying to study. For
Hodgkin–Huxley model by focusing on its bifurcations (Izhikevich,
example, map-based neurons are more suitable for large-scale
2003), although the equations resembles an Integrate-and-Fire
simulations or for generic studies (e.g. the cause of bifurcations,
(IF) model. Besides the behaviors presented in Fig. 8, this one also
the rise of chaotic orbits, synchronization dynamics, avalanche
presents: spike frequency adaptation (a behavior very similar to
dynamics, pattern recognition, memory modeling, etc.); and bio-
the one in the mixed mode panel in the aforementioned figure),
logically inspired models are more suitable for single cell studies
spike latency (a delayed response to a stimulus), depolarizing
(e.g. synaptic integration, dendritic cable filtering, effects of den-
after-potential (the membrane potential is slightly depolarized
dritic morphology, the dynamics of ionic currents, etc.) or even
right after the action potential), inhibitory induced tonic spiking
small-scale network simulations. Certainly, these suggestions hold
and inhibitory induced tonic bursting (the onset of tonic spiking or
only when one has to worry with the simulation times.
tonic bursting during the injection of an inhibitory direct current)
We have seen that maps have reduced computational time and
(Izhikevich and Hoppensteadt, 2004).
are a generic framework. Furthermore, regardless of the chosen
model, they bring other advantages:
3. Summary of the addressed models

So far we have presented a description of the earliest maps • There is no need for integration nor integration timestep adjust-
proposed to describe neuronal activity together with some recent ment; In fact, adjusting an integration timestep turns the ODE we
approaches. Izhikevich (2004) does an attempt to classify some are trying to solve into a discretized map. The bad thing in this
of these models, comparing them with the biologically motivated approach is that the stability of the ODE may not correspond to
models (HH-like neurons). To do so, he plots a graph with the the map stability.
amount of different biophysical behavior present in the model • The solution is exact; When solving an ODE, the adjustment of an
versus the amount of floating-point operations (FLOP – i.e. sums integration timestep turns the obtained solution in an approxi-
and multiplications) it would take for the computer to compute 1 mation.
ts of the model. • Maps are more plausible than pure cellular automata, also used to
The Hodgkin–Huxley model may produce any biophysically model neuronal activity (Copelli et al., 2002, 2005; Kinouchi and
plausible behavior, although it takes almost 100 times more FLOP. Copelli, 2006; Ribeiro et al., 2010). However, cellular automata
 M. Girardi-Schappo et al. / Journal of Neuroscience Methods 220 (2013) 116–130 123

Fig. 8. List of the KTz model excitable behaviors. The classification of these behaviors is based on Izhikevich and Hoppensteadt (2004). If not explicitly stated, the parameters
are K = 0.6, T = 0.35, ı = 0.001,  = 0.008, xR = −0.7 and initial conditions on the fixed point. External stimulus can be a delta pulse, I(t) = I0 ıt,t0 , a step current with intensity I0
or a current ramp, I(t) = a(t − t0 ). Interval between two stimuli is t. Time is in timesteps. (a) ı = 0.007,  = 0.004, I0 = 0.03; (b) I0 = 0.03; (c) ı = 0.004,  = 0.001, I0 = 0.04; (d)
 = 0.001, I0 = 0.04; (e) ı = 0.007,  = 0.003, I0 = 0.04; (f) a = 0.0008; (g) a = 0.0008; (h)  = 0.001, I0 = 0.4; (i) ı = 0.01, xR = −0.67, I0 = 0.05, t1 = 8, t2 = 232, t3 = 21; (j) ı = 0.004,
I0 = 0.06, t1 = 21, t2 = 219, t3 = 10; (k) I0 = −0.3; (l) ı = 0.004,  = 0.001, I0 = −0.3 during 50 timesteps; (m) ı = 0.004, I0 = ±0.06, t1 = 240, t2 = 14; (n) T = 0.33, ı = 0,  = 0,
z(0) = −0.01, I0 = 0.2, t = 228; (o) a = 0.0002 during 800 timesteps followed by a = 0.005 for 10 timesteps.

are very abstract entities in which not only the time, but also the are somewhat simpler, due to a smaller set of parameters, in
neuronal states are discretized. comparison with Rulkov or Izhikevich models.
• Maps keep the main dynamical and bifurcation properties of the
Hodgkin–Huxley-like neurons, as we have seen in the previous 4. Coupling maps
section.
The modeling of neurons as maps is an important and active
Concerning the choice of a particular map for simulations, we field of research in the past two decades, as we have seen in the
observe that there is a trade-off between computational efficiency previous sections of this paper. However, building up a network
and analytical tractability: the KTz map, and perhaps the Chialvo requires connecting the discrete time elements through synapses
map with the addition of a slow current to produce bursts, are distributed along a particular network topology. This section
not defined by piecewise functions, so their phase plane analysis is devoted to the synapse models. We summarize some types
124 M. Girardi-Schappo et al. / Journal of Neuroscience Methods 220 (2013) 116–130

0.0 The nonbiologically motivated models are usually rooted in the

Fast and paradigm of coupled map lattices (CML) (Kaneko, 1993; Kaneko and
Tsuda, 2001), which are not necessarily neuronal networks.
Slow Diffusive (electrical) synapses are fast connections, mediated
-0.2 Cardiac Spiking by gap junctions. They are usually modelled by Ohm’s law,
Spikes
where membrane potential differences generate synaptic currents

Bursting
through a fixed (or, sometimes, a time dependent) conductance.
-0.4
Impulsive (chemical) synapses are slow connections medi-
x R

ated by the exchange of neurotransmitters. They are thought to

Fixed be essential for neuronal communication, learning and memory
-0.6 phenomena (Keener and Sneyd, 1998). As in the case of neuronal
Point spikes, the postsynaptic potential produced by chemical synapses
may be modeled by biophysically realistic or by phenomenolo-
-0.8 gical dynamical equations. The equations may describe the time
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
T dependence of the conductances or of the postsynaptic currents.
In the next subsections, these three kind of models and their
applications will be briefly reviewed. The neurons can also con-
Fig. 9. Detailed phase diagram for the KTz model ( = ı = 0.001 and K = 0.6).
nect, disconnect, strengthen and weaken synapses as a function of
time. Although important, these issues will not be addressed in this
8 paper.

Bursting 4.1. Types of coupling

6
σth Some neuron models allow a very simple way of modelling
interactions, which will be described in the next subsection. These
α

interactions are generally used when studying CML and are not
biologically motivated.
1 Fixed However, all of the recently proposed neuron models provide
4
Point Spiking a specific way of inputting external currents into their equations.
Hence, we introduce here the letter Ii (t) as the total input current
over the ith neuron, which is, generally, given by the sum of the
synaptic currents,
2
-1 -0.5 0 0.5 1
σ 
Ni

Ii (t) = Yij (t), (18)

j=0
Fig. 10. Bifurcation diagram of the Rulkov model (Eq. (13)). The curve  th is given
by Eq. (15). The stability limit between the phases of bursting and spiking is not well where Yij (t) is the postsynaptic current from the presynaptic neuron
defined and has been determined computationally. Adapted from Rulkov (2002). j over postsynaptic neuron i at time t. Thus, modelling synapses for
such systems is equivalent to model the synaptic currents and post-
of connections, then we list some map models for them and synaptic potentials. These models are reviewed in the following
eventually discuss the main behaviors of these models. subsections.
In short, modeling synapses may or may not take into account
the types of connections biologically observed. We can classify the 4.1.1. Nonbiologically motivated coupling
simpler map-based synapse models in three categories: nonbiolog- A simplification of a neuronal network can be made if the neuron
ically motivated, biologically motivated with diffusive (electrical) model is disconnected from biological background and the cou-
character and biologically motivated with impulsive (chemical) plings between them are based in the paradigm of CML, i.e., the
character. functions are not necessarily rooted on biological foundations.

-54
1 Fixed Point

Bursting
-58
c

-62 Spiking

0 1 2 3 4
H
Fig. 11. Bifurcation diagram of the Izhikevich model (Eq. (16)) with a = 0.02, b = 0.25 and d = 0. The separation between spiking and bursting have been determined compu-
tationally and is not well defined, thus we traced a representative line. Adapted from Ibarz et al. (2011).
 M. Girardi-Schappo et al. / Journal of Neuroscience Methods 220 (2013) 116–130 125

The typical CML coupling can be written in the following gener- used to study the synchronization of bursts when the network is
alized mathematical form (Ibarz et al., 2011): not homogeneous (notice the indices in  i and in ˛i ). In this case,
pulse coupling allows chaotic bursts to synchronize.
 Ni
The next two kinds of coupling are biologically inspired and,
 (t + 1) = (1 − ˛)F(
X  X (t)) + ˛  X
F(  (t)), (19)
i i j hence, they are modeled by input currents just like the one used by
Ni
j=
/ i PCNN.

where X  (t) is the vector state of the ith neuron (node) at time t,
i 4.1.2. Diffusive coupling
with 1 ≤ i ≤ N, N being the total number of neurons, and Ni is the
Diffusive or electrical coupling is mediated by special chan-
total number of nodes connected to the ith node. ˛ ∈ [0 ; 1) is the
nels formed by connexin proteins (Connors and Long, 2004). These
coupling parameter: ˛ = 0 means no coupling and ˛ → 1 means that
passive channels between neighboring cells allow ions or small
the node’s Ni neighbors play a more important role to this element
molecules to pass through them. Their exchange speed allow faster
state than the node i itself.
signal propagation than chemical synapses.
Another kind of nonbiological coupling is to intertwine variables
The post synaptic current of electrical (Ohmic) synapses is
between elements of the network, like done by Güémez and Matías
modeled as (Connors and Long, 2004):
(1996) with two chaotic Chialvo neurons – one of them drives the
other – to study chaos suppression: 
Ni
 
Ii (t) = Gij xj (t) − xi (t) , (24)
2
x1 (t + 1) = [x1 (t)] exp[y1 (t) − x1 (t)] + I j=
/ i

y1 (t + 1) = ay1 (t) − bx1 (t) + c where xi (xj ) are the postsynaptic (presynaptic) membrane poten-
. (20)
tials and Gij is the total conductance of the gap junction. Here, the
x2 (t + 1) = [x1 (t)]2 exp[y2 (t) − x1 (t)] + I
synaptic current from neuron j to neuron i is Yij (t) = Gij [xj (t) − xi (t)].
y2 (t + 1) = ay2 (t) − bx2 (t) + c
4.1.3. Chemical synapses
Notice that x2 (t + 1) is function of x1 (t) and y2 (t), and not of x2 (t) The couplings via neurotransmitter (chemical synapses) are
as expected. Thus, it is said that the first neuron drives the second slower than electrical ones and the conductance is a nonlinear
one. The synchronization may suppress chaos in a couple of Chialvo function of time. The synaptic current Yij (t) is modeled as
neurons with intertwined variables.
It is important to notice that the framework of Eq. (19) does not Yij (t) = Gij (t)[xi (t) − Ej ] (25)
hold if one chooses to work with more elaborate neuron models,
where Gij (t) is the synaptic conductance, xi (t) is the membrane
like the KTz or the Rulkov models, because the phase plane anal-
potential of the postsynaptic neuron and Ej is the reversal synap-
ysis and the bifurcation diagrams of such models provide a very
tic potential. Notice that the value of Ej defines the excitatory or
specific way of inputing external currents into their equations, via
inhibitory character of the synapse.
the variable Ii (t).
The synaptic conductance may be modeled by (Roth and van
The so-called pulse-coupled neural network (PCNN) is an exam-
Rossum, 2010):
ple of simple coupling through input currents. The input over
neuron i is assumed to be: • an instantaneous rise with a single exponential decay:

Ni
t − tj
Ii (t) = Gij xj (t), (21) Gij (t) = [Gij (tj ) + G] exp (t − tj ), (26)

j=
/ i
where G is the instantaneous increment in Gij (t) when there is
where Gij is the coupling intensity (the conductance of the channel, a presynaptic firing of neuron j at time tj ,  is the characteristic
in the case of a synapse). The sum runs over the Ni neighbors of the decay time and (t) is the Heaviside function;
ith neuron and the xj (t)’s are the presynaptic membrane potentials. • an alpha function (with a continuous rise and fall):
Notice that Eq. (21) is a particular case of Eq. (18). It is called pulse t − tj t − tj
coupling because, as soon as the neuron j starts a pulse, it is readily Gij (t) = Gij exp − , (27)
 
transmitted to neuron i scaled by the intensity Gij in the next time
step. where Gij is the scaling parameter of the conductance;
As an example of PCNN, take the Kinouchi and Tragtenberg • a difference of exponentials:
(1996) approach to study emergence of a collective oscillatory  t − tj t − tj

state: a fully connected network of KT neurons (Section 2.2) Gij (t) = Gij exp − − exp − , (28)
1 2
coupled via Eq. (21) homogeneously (Gij = Gji = G). The so coupled
model reads: where Gij is the scaling and normalizing parameter of the conduc-
Ni tance and  1 and  2 are the characteristic rise and decay times of
xi (t) − Kyi (t) + H + G x (t)
/ i j
j= the synapse.
xi (t + 1) = tanh
T . (22)
Another approach is the fast threshold modulation (FTM)
yi (t + 1) = xi (t) approximation (Somers and Kopell, 1993), where there is a very
sharp activation threshold, , and a constant conductance, Gij , for
Another example is the chaotic Rulkov (2001) model in a mean
each coupling:
field pulse-coupled network, implemented by the equation:
Yij (t) = Gij [xi (t) − Ej ](xj (t) − ), (29)
G
N
˛
xi (t + 1) = 2
+ yi (t) + xj (t) where (x) is the step function.
1 + [xi (t)] N , (23)
j=
/ i Instead of using predetermined functions for Gij (t), we can adopt
yi (t + 1) = yi (t) − [xi (t) − ] a more coherent modeling, where synapses are also described
by map-based equations. For example, Rulkov et al. (2004) and
126 M. Girardi-Schappo et al. / Journal of Neuroscience Methods 220 (2013) 116–130

Fig. 13. Time evolution of the postsynaptic noisy current (bottom panel) of Eq. (31)
Fig. 12. Time evolution of the postsynaptic current, Eq. (31) for J = 0.0001,  1 =  2 = 15 with Jij (t) given by Eq. (32) with J = 0.001, R = 0.01 and  1 =  2 = 2 due to a presynaptic
(bottom panel) due to a presynaptic KTz neuron in bursting regime, Eq. (12) with KTz neuron spike (top panel). The KTz map parameters are K = 0.6, T = 0.35, ı = 0.001,
K = 0.6, T = 0.35, ı =  = 0.001 and xR = −0.5 (top panel).  = 0.008 and xR = −0.7.

Bazhenov et al. (2005) proposed a map-based model for the synap- linear sum of alpha or double-exponential functions (a common
tic current: procedure in computational neuroscience) nor by the other sim-
ple schemas described in the initial part of this section;
Yij (t + 1) = Yij (t) − Gij [xi (t) − Ej ]ıt,tj (30) • the chemical synapse is not always effective, then there is a prob-
where 0 ≤ < 1 is a decay parameter and the tj are the presynaptic ability q of exciting the postsynaptic neuron. Similar to the input
firing times and ıt,tj is the Kronecker delta. current threshold, above which a neuron fires, there is a threshold
Kuva et al. (2001) proposed previously a map-based chemical Jth , above which the postsynaptic neuron fires (the threshold Jth
coupling. The synaptic current is given by may also be negative when the postsynaptic neuron is excitable
by rebound).
1
Yij (t + 1) = 1− Yij (t) + Zij (t)
1
, (31) A first glance into Eq. (32) lead us to define the probability, p, of
1   having |Jij (t)| > |Jth |:
Zij (t + 1) = 1− Zij (t) + Jij  xj (t)
2
J + R − Jth
p= . (33)
where  1 and  2 are characteristic time constants, Jij is the coupling R
intensity (with dimension of current). This synaptic current may be Note that |J + R| > |Jth |, because J, R and Jth have always the same
excitatory (Jij > 0) or inhibitory (Jij < 0). Notice that one may choose sign. One could expect that q = p. That is not true because there is
to model the conductance dynamics by Eq. (31), replacing Yij (t) by an internal dynamics in the synaptic current equations: as long as
Gij (t) and taking Yij (t) = Gij (t)[xi (t) − Ej ]. the spike takes place (xj (t) > 0), the current keeps summing different
The solution of Eq. (31) is a discrete approximation to the dou- quantities J + ij (t). Thus, the same set (J, R) may not always lead to
ble exponential function, f(t) = C[exp(− t/ 1 ) − exp(− t/ 2 )], which a postsynaptic spike. Anyway, the quantity p may be regarded as a
reduces to an alpha function, f(t) = Ct exp(− t/), when  1 =  2 = ; C first approximation to q.
is a constant – see Fig. 12. Usually, these functions are used to model
the conductance of synaptic currents (see Eqs. (27) and (28)). How- 4.2. Applications of the synapse models
ever, it is still a good approximation for modeling synaptic currents
(de Schutter, 2010). A summary of the results obtained by each of the classes
One may question why there is no coupling parameter multi- of synapse models (nonbiologically motivated, electric, chemical)
plying the synaptic current, Yij (t). The coupling parameter is Jij . If in recognizing image patterns, modeling cat visual cortex, syn-
one wish to scale the current as Cij Yij (t), then a simple change of chronization (connected with several network topologies, noise,
variables in Eq. (31) may suppress Cij , keeping the same dynamical spike-burst behavior, delay, rewiring, etc.), bistable states of the
behavior: Yij (t) → Cij Yij (t), Zij (t) → Cij Zij (t) and Jij → Cij Jij . brain, effects of external signals, information transmission and crit-
Girardi-Schappo et al. (2012) studied KTz maps coupled by icality in neuronal avalanches is given in this subsection.
Kuva et al. synapses. The authors generalized the synaptic cur-
rent in Eq. (31) by adding a uniform random noise term in the 4.2.1. Nonbiologically motivated
coupling, in order to model synaptic noise observed experimen- As an example of CML application is the study of synchroniza-
tally (Katz and Miledi, 1967; London et al., 2002). Then, the tion made by Jampa et al. (2007). They connected Chialvo neurons
Girardi–Schappo–Kinouchi–Tragtenberg (GKT) model uses Eq. (31) through a short version of Eq. (19), in which the coupling term is
with Jij ≡ Jij (t) given by: written only as ( /2)[xj(t) (t) + xk(t) (t)], where the indices j(t) and k(t)
are picked at each timestep, with probability p, by randomly assign-
Jij (t) = J + ij (t), (32)
ing to each node an index, l ∈ [1 ; N], and then calculating j(t) = l − 1
where J is the coupling parameter and ij (t) is a uniform random and k(t) = l + 1 (modulo N). The rewiring probability p and the cou-
time signal of amplitude |R|. The noise is independent for every pling strength determine the kind of synchronization: chaotic,
synapse j → i. To keep the synaptic coherence (i.e. in order to the spatio-temporal chaos or fixed point.
keep inhibitory synapses always inhibiting and excitatory synapses Pontes et al. (2008) studied a CML of Rulkov neurons with
always exciting), the sign of R is the same sign of J, then if J > 0, we power-law decreasing coupling intensity within a range neighbors
have ij (t) ∈ [0, R], otherwise we have ij (t) ∈ [R, 0]. A realization of of each neuron. They found that the coupling strength needed to
the noisy synaptic current is plotted in Fig. 13. synchronize chaotic Rulkov bursters is smaller when the range of
Two important features are captured by this model: couplings is larger.
de Vries (2001a,b) showed that mean field-coupling of the same
• time summation of postsynaptic currents is nonlinear (see system can turn isolated spiking into coupled bursting neurons.
Fig. 12), an empirical fact that cannot be modeled by a simple Then, in this case bursting can be seen as an emergent behavior.
 M. Girardi-Schappo et al. / Journal of Neuroscience Methods 220 (2013) 116–130 127

Luccioli et al. (2012) used coupled logistic maps, as well as different polychronization patterns. Yet, the number of produced
Stuart–Landau oscillators and leaky IF neurons, to study dynam- patterns is far greater than the number of neurons in the network,
ical properties of coupled oscillator sparse lattices in many scales yielding a very large memory capacity.
of length. They show a transition from micro to macroscopic scale
in network synchronization driven by a critical connectivity.
4.2.2. Diffusive coupling
López-Ruiz and Fournier-Prunaret (2012) mimicked a bistable
Diffusive coupling has been extensively used in the study of
(sleep-awaken) brain by a logistic map lattice, with usual CML inter-
synchronization, spatio-temporal chaos, effects of external inputs
actions.
on a neuronal network, spiking-bursting transition, among other
As examples of the PCNN approach, we list:
subjects.
Rulkov (2002) studied the existence of synchronization regimes
• Eckhorn et al. (1990) modeled the cat visual cortex, particularly
for spiking and bursting activities of the Rulkov map as a function
the stimulus-induced oscillations, and studied pattern recogni- of neuron coupling strength.
tion and feature associations; Tanaka et al. (2006) checked the influence of a mix of electrical
• Kinouchi and Tragtenberg (1996) showed that pulse coupled
and chemical synapses connecting Rulkov neurons on burst syn-
identical KT neurons synchronize as one single KT neuron with chronization and transmission (interpreted as chaotic itinerancy
rescaled parameters; (Kaneko and Tsuda, 2003)).
• Goel and Ermentrout (2002) discretized an ODE, and found waves
Wang et al. (2007) found that subthreshold stimulus induce
in a ring network and synchronization stability dependent of the synchronization in a noisy square lattice of Rulkov maps. Wang
number of neurons; et al. (2008) studied Rulkov neurons with information transmission
• Izhikevich (2003) studied brain rhythms with pulse coupled
delay in a small-world geometry with additive noise, and showed
maps given by the Izhikevich model (discretized with a 0.1 ms the existence of transitions in synchronization as a function of the
timestep); delay. The delay may also induce stochastic resonance in a scale-
• Zou et al. (2009) found stable chaos in a PCNN of discontinuous
free network of Rulkov neurons (Wang et al., 2009).
maps; Chen et al. (2008) found that there is an optimal range of het-
• Wang et al. (2010) reviewed the successful use of PCNN in many
erogeneity in the  e coupling parameter of the Rulkov neurons that
dimensions of image processing: segmentation, denoising, object may result in coherence resonance.
and edge detection, feature extraction and pattern recognition Ivanchenko et al. (2004, 2007) studied transitions between
(enhancement, fusion, etc.); bursting and spiking Rulkov neurons connected by electrical
• Lameu et al. (2012) coupled chaotic Rulkov neurons in a clus-
synapses in many time scales. The authors also studied the chaotic
tered scale-free architecture and studied means of suppressing phase synchronization by an external periodic input to a single
bursting synchronization, motivated by biological reasons. neuron.
Batista et al. (2007, 2009) also showed that external periodic sig-
Specifically, Kinouchi and Tragtenberg (1996) showed that the nals can generate burst synchronization in networks of electrically
all-to-all pulse-coupled KT network with homogeneous couplings, coupled Rulkov neurons.
Gij = G in Eq. (21), may lead a network of silent neurons (adjusted Agnes et al. (2010) studied synchronization of Rulkov neurons
with (K, T, H < Hs )) into a collective oscillating phase if the sum Ii (t) =
N coupled homogeneously with gap junctions in a square lattice and
G j =/ i xj (t) is sufficiently large, such that H̃ = H + Ii (t) > Hs (K, T ) explored the conditions under which collective states may show
(for neurons with H < 0). up in the dynamics of the network.
More clearly, if H < 0, the fixed points are also xi∗ < 0. Thus, if Luccioli et al. (2012) connected Rulkov neurons in a small-world
N network through electrical connections to study synchronization
G < 0 (inhibitory coupling), the sum Ii (t) = G j =/ i xj∗ will be a pos-
and spatio-temporal chaos.
itive number and the achieved state, after the transient time, will
have a new H̃ given by H̃ = H + Ii (t). Then, there is a threshold –
exactly equal to the excitability threshold discussed in Section 2.2 – 4.2.3. Impulsive coupling
above which the network will spontaneously oscillate, even though Connecting map neurons by chemical synapses is not as pop-
any uncoupled neuron is silent. This is a counter-intuitive result, ular as using electrical synapses, even though chemical synapses
although it also happens with realistic neurons (Golomb and Rinzel, are more commonly found in the brain. One of the most popular
1993; Ernst et al., 1995). approaches is the FTM, which presents no synaptic dynamics. How-
If the term |G| is large enough, the network crosses through the ever, there are studies that use dynamical (map-based) synapses.
bulb in Fig. 2 and reaches the upper fixed point, x* > 0. This picture Ivanchenko et al. (2007) showed that networks of Rulkov maps
may be generalized to the heterogeneous case, however the value coupled with electrical and FTM chemical synapses exhibit a
of Ii (t), in such case, is not guaranteed to be positive. synchronized–desynchronized transition as a function of the cou-
The heterogeneous case may be used both for modeling practical pling strength.
problems and for investigation purposes. As examples, one could Ibarz et al. (2008) used FTM to show many effects produced by
assemble a neural network for time series learning by mixing lin- impulsive synapses: excitatory synapses may generate antiphase
ear neurons (in the output) and KT units in the processing layers synchronization, the changing character of synapses from excit-
and adjusting weights according to a training algorithm (Kinouchi atory to inhibitory when the conductance changes and the reversal
and Tragtenberg, 1996). See also Albano et al. (1992) for an applica- potential keeps the same, and small variations in the synaptic
tion with neurons given by Eq. (2), but with a nonlinear activation threshold may cause dramatic changes in synchronization of spikes
function. within bursts.
The Izhikevich heterogeneous network consists in mixed Shi and Lu (2009) also applied FTM coupling in the study of in-
excitatory (Gij > 0) and inhibitory (Gij < 0) coupling. From the phase burst synchronization of Rulkov neurons.
network’s time evolution, groups of synchronized neurons emerge Franovic and Miljkovic (2010) showed that bursting Rulkov
in different phases – the so called polychronous state. Izhikevich maps coupled with inhibitory chemical synapses may exhibit
(2006) proposes that these polychronous groups represent the power-law behavior at the onset of a synchronization transition,
memory of the network, as different external stimuli produce driven by the coupling strength or stimulation current.
128 M. Girardi-Schappo et al. / Journal of Neuroscience Methods 220 (2013) 116–130

Franovic and Miljkovic (2011) studied the effects of synaptic criticality in the brain (Chialvo, 2010; Werner, 2010; Beggs and
delay in the synchronization of networks of chaotic Rulkov neu- Timme, 2012; Shew and Plenz, 2013).
rons connected by reciprocal chemical synapses. The synapses The basic idea is that the brain works with the precisely needed
are modeled by summing in each neuron the standard analytical amount of activity, so that signals neither decay exponentially in
expression for synaptic conductance. Except that, in the referred space nor saturate the network, and that it may exists a homeostatic
work, synapses’ threshold is modeled by a sigmoid function, allow- mechanism which leads the critical tissues to such state. Within this
ing the authors to study graded synaptic transmission and fast framework, some mental diseases and disorders would be associ-
threshold modulation. ated with deviations from the critical activity (Chialvo, 2004, 2010;
Batista et al. (2012) used a modified FTM coupling (in which the Vertes et al., 2011).
Heaviside function is replaced by a sigmoid function) to study con- The critical state has been shown to enhance the dynamical
ditions under which there are synchronized regimes in a clustered range of sensitivity to external stimuli (Kinouchi and Copelli, 2006;
small-world network of chaotic Rulkov neurons. Shew et al., 2009; Larremore et al., 2011), to optimize the memory
Rulkov et al. (2004) designed cortical networks consisting of reg- and learning processes (de Arcangelis et al., 2006) and to enhance
ular spiking neurons and fast spiking interneurons to model sleep the computational power of the brain (Werner, 2010; Shew and
and activated states of the thalamocortical system. The authors dis- Plenz, 2013; Beggs and Timme, 2012).
cuss that map-based models can be widely used for large-scale A necessary condition for SOC is that the system may not be
simulations and that such models are especially useful for tasks imposed into the critical state, rather it should find its own path
where the modeling of specific firing patterns of different cell towards criticality with no fine tuning. Models with dynamical
classes is important. synapses lead to critical or quasicritical networks (Levina et al.,
Using basically the same model, Bazhenov et al. (2005) deter- 2007). In this sense, the GKT model is not SOC, because the authors
mined resonance properties on collective behavior in a cortical had to adjust the parameters J and R. However, a variation of the GKT
network model containing excitatory and inhibitory cells and model could contain a homeostatic mechanism over the parameter
showed that network interactions can enhance the frequency range R or J, for instance, that would take the system towards the critical
of reliable responses, such that the latter can be controlled by the line, turning the system into a SOC system (Girardi-Schappo et al.,
strength of synaptic connections. in preparation).
Agnes et al. (2012) modified Eq. (30) in order to account for
Hebbian’s rules of learning and coupled chaotic Rulkov neurons 5. Concluding remarks
in a square lattice. They studied the dynamical properties and the
fixed points of the resulting system. Although generally studied in the context of neuronal dynam-
Girardi-Schappo et al. (2012), using the Kuva et al. (2001) ical systems, some of the models outlined in this work may
noisy synapse model, could study the effects of the synaptic effi- be used in computational applications, such as pattern recog-
cacy on a network of KTz neurons. The noise turns the problem nition, data analysis, data classification, data association and so
into a matter of stochastic dynamics, since the amount of neu- on, like the McCulloch–Pitts model utilized in the perceptron.
rons that fires in the network due to an initial stimulation is not We directed our review mainly through the historical devel-
obvious. opment of the map-based models, by theoretically constructing
The network activity due to a single stimulus is called an over the McCulloch–Pitts model, adding delayed self-couplings
avalanche. The authors investigated regular square lattices with (which creates new dynamical behaviors), changing to a con-
free boundary conditions of KTz neurons coupled with GKT tinuous activation function and adding synaptic currents to the
synapses. They found a phase transition line in which the number model.
of spikes in an avalanche, s, follows a cutoff power-law P(s) ∼ s− This approach allowed us to close the gap between the first dis-
with  = 1.35. The duration of avalanches, w, also follows a cutoff crete time models and the most recent maps. In general, we tried
power-law P(w)∼w− , with = 1.50. to keep the original concepts of the models, except for the Izhike-
Moreover, the authors measured the activity of a subsampled vich’s, which was originally designed as an ODE. Nevertheless we
network – a network in which the measurement can be made in assume that the discretization of an ODE results in a map, although,
only a very small fraction of the elements. Generally, experimen- sometimes, the resulting map may not provide a dynamically sta-
tal setups are subsampled, since the number of neurons which ble approximation of the initial ODE behavior if the time step is
are recorded is far less than the number of neurons in the sample large. Thus, we linked two entire family of models with very rich
(Priesemann et al., 2009; Ribeiro et al., 2010). excitable and dynamical behaviors, namely the perceptron fam-
The subsampling results are in accordance with the experi- ily (from McCulloch–Pitts to KTz) and the Rulkov family (Chialvo,
ments of Ribeiro et al. (2010) and, together with the power-law Rulkov, Courbage et al. and Izhikevich). We also showed that there
behavior in the network activity – both in temporal and spatial is a close dynamical correspondence between KTz and the behav-
dimensions – they show that the system is critical and support iors of conductance based models.
the fact that there may be a self-organized critical (SOC) state in The models of the Rulkov family are, generally, piecewise
networks of dynamically realistic neurons (Girardi-Schappo et al., defined functions. All of these models present bursting activity with
2012). only two dynamical variables, differently from the KTz map (whose
The SOC state is, generally, present when there is a balance bursting mechanism is based on the Hindmarsh–Rose model). On
between the slow accumulation of tension and the fast relaxation the other hand, the symmetry and simplicity of the KTz map may
of the system (Bak et al., 1987, 1988; Christensen et al., 1993; be of great help during the phase plane and bifurcation analysis
Jensen, 1998; Dhar, 2006; Bonachela and Muñoz, 2009, and many (recall the many phase diagrams presented for this map, whereas
others). Due to theoretical considerations about criticality in there are only few phase diagrams for the Rulkov family).
neural networks (Usher et al., 1995; Stassinopoulos and Bak, 1995; KT and KTz may also be suitable when there is a need to explore
Herz and Hopfield, 1995), Beggs and Plenz (2003) proposed SOC the effects of the whole action potential, as all the other mod-
to explain spontaneous activity in cultured hippocampal cells. els, including the integrate-and-fire models, assume that the spike
Neuronal avalanches are experimentally observed (Beggs and occurs instantly. Therefore, if one wants to model a more rigorous
Plenz, 2003, 2004; Priesemann et al., 2009; Ribeiro et al., 2010, synaptic transmission with the map framework, KT or KTz could be
and references there in), although there is an ongoing debate on a good compromise.
 M. Girardi-Schappo et al. / Journal of Neuroscience Methods 220 (2013) 116–130 129

Every map model presents a similar computational perfor- Batista CAS, Batista AM, de Pontes JAC, Lopes SR, Viana RL. Bursting synchronization
mance, as discussed in Section 3, which is approximately twenty in scale-free networks. Chaos Solitons Fractals 2009];41:2220–5.
Batista CAS, Batista AM, de Pontes JAC, Viana RL, Lopes SR. Bursting synchronization
times faster than the HH performance. Notice that the standard in scale-free networks. Physical Review E 2007];76:016218.
HH model has only two voltage-dependent conductances, namely Batista CAS, Lameu EL, Lopes AMBSR, Pereira T, Zamora-Lópes G, Kurths J, et al.
sodium and potassium. Adding other conductances (calcium, chlo- Phase synchronization of bursting neurons in clustered small-world networks.
Physical Review E 2012];86:016211.
rine, etc.) implies in slower computational times. Thus, the most Bazhenov M, Rulkov NF, Fellous J, Timofeev I. Role of network dynamics in shaping
prominent features of all the studied neuronal maps are their spike timing reliability. Physical Review E 2005];72:041903.
simplicity, reliability, numerical stability and computational per- Beggs JM, Plenz D. Neuronal avalanches in neocortical circuits. Journal of Neuro-
science 2003];23(35):11167–77.
formance.
Beggs JM, Plenz D. Neuronal avalanches are diverse and precise activity patterns
Moreover, we classified the many types of coupling used in that are stable for many hours in cortical slice cultures. Journal of Neuroscience
the literature to connect the referred models. Hence, we tried to 2004];24(22):5216–29.
Beggs JM, Timme N. Being critical of criticality in the brain. Frontiers in Physiology
standardize the notation, instead of following the original ones
2012];3:163.
in order to highlight the similarities and differences between the Bellman RE. Dynamic Programming. Dover; 2003].
proposed coupling schemes. Many times, the coupling and the net- Bonachela JA, Mu noz MA. Self-organization without conservation: true or just
work topology are indissociable, forcing one to study the coupling apparent scale-invariance? Journal of Statistical Mechanics 2009]:P09009.
Bower JM, Beeman D. The book of GENESIS: exploring realistic neural models with
of neurons under a given topology. From pulse-coupled networks the GEneral NEural SImulation System. Internet Edition; 2003].
to map-based modeling of chemical synapses, we presented their Caianello ER. Outline of a theory of thought process and thinking machines. Journal
most common implementations, relating with the results achieved of Theoretical Biology 1961];1:204–35.
Carnevale NT, Hines ML. The NEURON book. Cambridge University Press; 2006].
in each case. Chen H, Zhang J, Liu J. Enhancement of neuronal coherence by diversity in coupled
The dynamics of synapses is a broad research area. We high- rulkov-map models. Physica A 2008];387:1071–6.
lighted the importance of modeling synapses specifically for Chialvo DR. Generic excitable dynamics on a two-dimensional map. Chaos Solitons
Fractals 1995];5:461–79.
working with map-based neural networks, as did Kuva et al. Chialvo DR. Critical brain networks. Physica A 2004];340:756–65.
(2001). Important synapse features, as noise and plasticity, can be Chialvo DR. Emergent complex neural dynamics. Nature Physics 2010];6:
introduced directly in these types of equations (Girardi-Schappo 744–50.
Christensen K, Flyvbjerg H, Olami Z. Self-organized critical forest-fire model: mean-
et al., in preparation).
field theory and simulation results in 1 to 6 dimensions. Physical Review Letters
Finally, recent reviews have shown that map-based neural 1993];71(17):2737–40.
networks is a promising developing field (Aihara and Suzuki, 2010; Cody WJ, Waite W. Software manual for the elementary functions. Prentice-Hall;
1980].
Ibarz et al., 2011). Just as examples, on the list of some still
Connors BW, Long MA. Electrical synapses in the mammalian brain. Annual Review
unexplored paths in map-based neuronal modeling are the com- of Neuroscience 2004];27:393–418.
partmental modeling (using electrically coupled maps to describe Copelli M, Oliveira RF, Roque AC, Kinouchi O. Signal compression in the sensory
compartments, like dendritic branchlets and axons) and the seek periphery. Neurocomputing 2005];65–66:691–6.
Copelli M, Roque AC, Oliveira RF, Kinouchi O. Physics of psychophysics: stevens and
for a canonical purely map-based model. A canonical map-based weber-fechner laws are transfer functions of excitable media. Physical Review
model would be the simplest model capable of describing all the E 2002];65:060901.
HH-type neurons behaviors (Hoppensteadt and Izhikevich, 2002). Copelli M, Tragtenberg MHR, Kinouchi O. Stability diagrams for bursting neurons
modeled by three-variable maps. Physica A 2004];342:263–9.
Another important issue is to develop schemas for short-term and Courbage M, Nekorkin VI. Map based models in neurodynamics. International Jour-
long-term synaptic plasticity (Kinouchi and Kinouchi, 2010). nal of Bifurcation and Chaos 2009];20(6):1631–51.
With this work, we hope we have brought more attention to Courbage M, Nekorkin VI, Vdovin LV. Chaotic oscillations in a map-based model of
neural activity. Chaos 2007];17(4):043109.
this kind of modeling, which may play an important role in the Cross MC, Hohenberg PC. Pattern formation outside of equilibrium. Reviews of Mod-
forthcoming years, both in technologic applications and in neuro- ern Physics 1993];65:851–1112.
scientific research as a new computational neuroscience method. Cymbalyuk G, Calabrese RL. Oscillatory behaviors in pharmacologically iso-
lated heart interneurons from the medicinal leech. Neurocomputing
2000];32–33:97–104.
Dayan P, Abbott LF. Theoretical neuroscience: computational and mathematical
Acknowledgements modeling of neural systems. The MIT Press; 2001].
de Arcangelis L, Perrone-Capano C, Herrmann HJ. Self-organized criticality model
for brain plasticity. Physical Review Letters 2006];96:028107.
We would like to thank the invitation to make this neuron
de Schutter E, editor. Computational modeling methods for neurocientists. The MIT
map modeling overview from Antonio Carlos Roque da Silva. MGS Press; 2010].
acknowledges partial financial support from CAPES and CNPq and de Vries G. Bursting as an emergent phenomenon in coupled chaotic maps. Physical
OK thanks the funding from CNAIPS-USP. Review E 2001a];64:051914.
de Vries G. From spikers to bursters via coupling: help from heterogeneity. Bulletin
of Mathematical Biology 2001b];63:371–91.
Dhar D. Theoretical studies of self-organized criticality. Physica A 2006];369:29–70.
References Eckhorn R, Reitboeck HJ, Arndt M, Dicke PW. Feature linking via synchronization
among distributed assemblies: simulations of results on cat visual cortex. Neural
Agnes EJ, Erichsen-Jr R, Brunnet LG. Synchronization regimes in a map-based model Computation 1990];2:293–307.
neural network. Physica A 2010];389:651–8. Ernst U, Pawelzik K, Geisel T. Synchronization induced by temporal delays in pulse-
Agnes EJ, Erichsen-Jr R, Brunnet LG. Model architecture for associative memory in a coupled oscillators. Physical Review Letters 1995];74:1570–3.
neural network of spiking neurons. Physica A 2012];391(3):843–8. FitzHugh R. Mathematical models of threshold phenomena in the nerve membrane.
Aihara K, Suzuki H. Theory of hybrid dynamical systems and its applications to bio- Bulletin of Mathematical Biophysics 1955];17:257–78.
logical and medical systems. Philosophical Transactions of the Royal Society A Franovic I, Miljkovic V. Power law behavior related to mutual synchroniza-
2010];368:4893–914. tion of chemically coupled map neurons. European Physical Journal B
Aihara K, Takabe T, Toyoda M. Chaotic neural networks. Physics Letters A 2010];76:613–24.
1990];144(6,7):333–40. Franovic I, Miljkovic V. The effects of synaptic time delay on motifs of chemi-
Albano AM, Passamante A, Hediger T, Farrell ME. Using neural nets to look for chaos. cally coupled rulkov model neurons. Communications in Nonlinear Science and
Physica D 1992];58:1–9. Numerical Simulation 2011];16:623–33.
Azevedo FAC, Carvalho LRB, Grinberg LT, Farfel JM, Ferretti REL, Leite REP, et al. Girardi-Schappo M, Kinouchi O, Tragtenberg MHR. Critical avalanches and subsam-
Equal numbers of neuronal and nonneuronal cells make the human brain pling in map-based neural networks; 2012. arXiv:1209.3271 [cond-mat.dis-nn].
an isometrically scaled-up primate brain. Journal of Comparative Neurology Girardi-Schappo M, Kinouchi O, Tragtenberg MHR. Neuronal avalanches due to self-
2009];513:532–41. organized noisy synapses; 2013 (in preparation).
Bak P, Tang C, Wiesenfeld K. Self-organized criticality: an explanation of 1/f noise. Goel P, Ermentrout B. Synchrony, stability, and firing patterns in pulse-coupled
Physical Review Letters 1987];59(4):381–4. oscillators. Physica D 2002];163:191–216.
Bak P, Tang C, Wiesenfeld K. Self-organized criticality. Physical Review A Golomb D, Rinzel J. Dynamics of globally coupled inhibitory neurons with hetero-
1988];38(1):364–74. geneity. Physical Review E 1993];48:4810–4.
130 M. Girardi-Schappo et al. / Journal of Neuroscience Methods 220 (2013) 116–130

Gua H-G, Jia B, Lia Y-Y, Chen G-R. White noise-induced spiral waves and multiple Luccioli S, Olmi S, Politi A, Torcini A. Collective dynamics in sparse networks. Physical
spatial coherence resonances in a neuronal network with type i excitability. Review Letters 2012];109:138103.
Physica A 2013];392:1361–74. McCulloch WS, Pitts WH. A logical calculus of the ideas immanent in nervous activity.
Güémez J, Matías MA. Synchronous oscillatory activity in assemblies of chaotic Bulletin of Mathematical Biophysics 1943];5:115–33.
model neuron. Physica D 1996];96:334–43. Morris C, Lecar H. Voltage oscillations in the barnacle giant muscle fiber. Biophysics
Herz AVM, Hopfield JJ. Earthquake cycles and neural reverberations: collective oscil- Journal 1981];35:193–213.
lations in systems with pulse-coupled threshold elements. Physical Review Nagumo J, Arimoto S, Yoshizawa S. An active pulse transmission line simulating
Letters 1995];75:1222–5. nerve axon. Proceedings of the IRE 1962];50:2061–70.
Hindmarsh JL, Rose RM. A model of neuronal bursting using three coupled first order Nagumo J, Sato S. On a response characteristic of a mathematical neuron model.
differential equations. Proceedings of the Royal Society B: Biological Sciences Kybernetik 1972];10:155–64.
1984];221:87–102. Pasemann F. Dynamics of a single model neuron. International Journal of Bifurcation
Hopfield JJ. Neurons with graded response have collective computational properties and Chaos 1993];2:271–8.
like those of two-sate neurons. Proceedings of the National Academy of Sciences Pasemann F. A simple chaotic neuron. Physica D 1997];104:205–11.
of the United States of America 1984];81:3088–92. Pontes JCA, Viana RL, Lopes SR, Batista CAS, Batista AM. Bursting synchronization in
Hoppensteadt FC, Izhikevich EM. Canonical neural models. The MIT Press; 2002]. non-locally coupled maps. Physica A 2008];387:4417–28.
Horn D. Networks of complex neurons. Physica A 1993];200:594–601. Priesemann V, Munk MHJ, Wibral M. Subsampling effects in neuronal avalanche
Ibarz B, Cao H, Sanjuan MAF. Bursting regimes in map-based neuron mod- distributions recorded in vivo. BMC Neuroscience 2009];10:40.
els coupled through fast threshold modulation. Physical Review E 2008];77: Ribeiro TL, Copelli M, Caixeta F, Belchior H, Chialvo DR, Nicolelis MAL, et al. Spike
051918. avalanches exhibit universal dynamics across the sleep–wake cycle. PLoS ONE
Ibarz B, Casado JM, Sanjuán MAF. Map-based models in neuronal dynamics. Physics 2010];5(11):e14129.
Reports 2011];501:1–74. Roth A, van Rossum MCW. Modeling synapses. The MIT Press; 2010].
Ivanchenko MV, Osipov GV, Shalfeev VD, Kurths J. Phase synchronization in ensem- Rulkov NF. Regularization of synchronized chaotic bursts. Physical Review Letters
bles of bursting oscillators. Physical Review Letters 2004];93:134101. 2001];86:183–6.
Ivanchenko MV, Osipov GV, Shalfeev VD, Kurths J. Network mechanism for burst Rulkov NF. Modeling of spiking-bursting neural behavior using two-dimensional
generation. Physical Review Letters 2007];98:108101. map. Physical Review E 2002];65:041922.
Izhikevich EM. Simple model of spiking neurons. IEEE Transactions on Neural Rulkov NF, Timofeev I, Bazhenov M. Oscillations in large-scale cortical networks:
Networks 2003];14(6):1569–72. map-based model. Journal of Computational Neuroscience 2004];17:203–23.
Izhikevich EM. Which model to use for cortical spiking neurons? IEEE Transactions Shew WL, Plenz D. The functional benefits of criticality in the cortex. Neuroscientist
on Neural Networks 2004];15:1063–70. 2013];19(1):88–100.
Izhikevich EM. Polychronization: computation with spikes. Neural Computation Shew WL, Yang H, Petermann T, Roy R, Plenz D. Neuronal avalanches imply maxi-
2006];18(2):245–82. mum dynamic range in cortical networks at criticality. Journal of Neuroscience
Izhikevich EM, Hoppensteadt F. Classification of bursting mappings. International 2009];29(49):15595–600.
Journal of Bifurcation and Chaos 2004];14(11):3847–54. Shi X, Lu Q. Burst synchronization of electrically and chemically coupled map neu-
Jampa MPK, Sonawane AR, Gade PM, Sinha S. Sincronization in a network of model rons. Physica A 2009];388:2410–9.
neurons. Physical Review E 2007];75:026215. Shilnikov AL, Rulkov NF. Subthreshold oscillations in a map-based neuron model.
Jensen HJ. Self-organized criticality: emergent complex behavior in phyical and Physics Letters A 2004];328:177–84.
biological systems. Cambridge University Press; 1998]. Somers D, Kopell N. Rapid synchronization through fast threshold modulation. Bio-
Kaneko K. Theory and applications of coupled map lattices. Wiley; 1993]. logical Cybernetics 1993];68:393–407.
Kaneko K. Relevance of dynamic clustering to biological networks. Physica D Stassinopoulos D, Bak P. Democratic reinforcement: a principle for brain function.
1994];75:55–73. Physical Review E 1995];51(5):5033–9.
Kaneko K, Tsuda I. Complex systems: chaos and beyond. A constructive approach Tanaka G, Ibarz B, Sanjuán MAF, Aihara K. Synchronization and propagation of bursts
with applications in life sciences. Springer; 2001]. in networks of coupled map neurons. Chaos 2006];16:013113.
Kaneko K, Tsuda I. Chaotic itinerancy. Chaos 2003];13:926–36. Tragtenberg MHR, Yokoi CSO. Field behavior of an ising model with competing
Katz B, Miledi R. A study of synaptic transmission in the absence of nerve impulses. interactions on the bethe lattice. Physical Review E 1995];52(3):2187–97.
Journal of Physiology 1967];192:407–36. Usher M, Stemmler M, Olami Z. Dynamic pattern formation leads to 1/f noise in
Keener J, Sneyd J. Mathematical physiology. Springer; 1998]. neural populations. Physical Review Letters 1995];74:326–9.
Kinouchi O, Copelli M. Optimal dynamical range of excitable networks at criticality. Veiga FLS, Tragtenberg MHR. A very stochastic resonant neuron model. Neurocom-
Nature Physics 2006];2:348–51. puting 2001];38–40:423–31.
Kinouchi O, Kinouchi RR. Dreams, endocannabinoids and itinerant dynamics in Vertes PE, Bassett DS, Duke T. Scale-free statistics of neuronal assemblies predict
neural networks: re elaborating crick-mitchison unlearning hypothesis; 2010. learning performance. BMC Neuroscience 2011];12(Suppl 1):O4.
arXiv:cond-mat/0208590 [cond-mat.dis-nn]. Wang QY, Duan Z, Perc M, Chen GR. Synchonization transitions on small-world
Kinouchi O, Tragtenberg MHR. Modeling neurons by simple maps. International neuronal networks: effects of information transmission delay and rewiring
Journal of Bifurcation and Chaos 1996];6:2343–60. probability. Europhysics Letters 2008];83:50008.
Kuva SM, Lima GF, Kinouchi O, Tragtenberg MHR, Roque AC. A minimal Wang QY, Lu QS, Chen GR. Subthreshold stimulus-aided temporal order and syn-
model for excitable and bursting elements. Neurocomputing 2001];38–40: chronization in a square lattice noisy neuronal network. Europhysics Letters
255–61. 2007];77:10004.
Lameu EL, Batista CAS, Batista AM, Iarosz K, Viana RL, Lopes SR, et al. Suppression of Wang QY, Perc M, Duan Z, Chen GR. Delay-induced multiple stochastic resonances
bursting synchronization in clustered scale-free (rich-club) neuronal networks. on scale-free neuronal networks. Chaos 2009];19:023112.
Chaos 2012];22:043149. Wang Z, Ma Y, Cheng F, Yang L. Review of pulse-coupled neural networks. Image
Larremore DB, Shew WL, Restrepo JG. Predicting criticality and dynamic and Vision Computing 2010];28:5–13.
range in complex networks: effects of topology. Physical Review Letters Werner G. Fractals in the nervous system: conceptual implications for theoretical
2011];106:058101. neuroscience. Frontiers in Physiology 2010];1:15.
Levina A, Herrmann JM, Geisel T. Dynamical synapses causing self-organized criti- Wua X, Maa J, Lia F, Jia Y. Development of spiral wave in a regular network of excit-
cality in neural networks. Nature Physics 2007];3:857–60. atory neurons due to stochastic poisoning of ion channels. Communications in
Little WA. The existence of persistent states in the brain. Mathematical Biosciences Nonlinear Science and Numerical Simulation 2013];18(12):3350–64.
1974];19:101–20. Yokoi CSO, de Oliveira MJ, Salinas SR. Strange attractor in the ising model
London M, Schreibman A, Häusser M, Larkum ME, Segev I. The information efficacy with competing interactions on the cayley tree. Physical Review Letters
of a synapse. Nature Neuroscience 2002];5:332–40. 1985];54(3):163–6.
López-Ruiz R, Fournier-Prunaret D. The bistable brain: a neuronal model with sym- Zou H, Guan S, Lai C-H. Dynamical formation of stable irregular transients in discon-
biotic interactions; August 2012. arXiv:nlin.CD/1208.0223v1. tinuous map systems. Physical Review E 2009];80:046214.