On a phase diagram for random neural networks with embedded spike timing dependent plasticity

BioSystems 89 (2007) 280 286 On a phase diagram for random neural networks with embedded spike timing dependent plasticity Tatyana S. Turova a,, Alessandro E.P. Villa b,c,1 a Mathematical Center, Lund University, Sweden b Laboratoire de Neurobiophysique, University Joseph Fourier, Inserm, U318 Grenoble, France c Neuroheuristic Research Group, Institute of Computer Science and Organization INFORGE, University of Lausanne, Switzerland Received 28 November 2005; accepted 26 May 2006 Abstract This paper presents an original mathematical framework based on graph theory which is a first attempt to investigate the dynamics of a model of neural networks with embedded spike timing dependent plasticity. The neurons correspond to integrate-and-fire units located at the vertices of a finite subset of 2D lattice. There are two types of vertices, corresponding to the inhibitory and the excitatory neurons. The edges are directed and labelled by the discrete values of the synaptic strength. We assume that there is an initial firing pattern corresponding to a subset of units that generate a spike. The number of activated externally vertices is a small fraction of the entire network. The model presented here describes how such pattern propagates throughout the network as a random walk on graph. Several results are compared with computational simulations and new data are presented for identifying critical parameters of the model. 2006 Elsevier Ireland Ltd. All rights reserved. Keywords: Random network; Spike timing dependent synaptic plasticity; Spiking neural network; Graph theory 1. Introduction The dynamics of connectivity patterns within large networks of integrate-and-fire neuromimes with embedded spike timing dependent plasticity (STDP) rules represents one of the most intriguing questions under investigation in recent years. STDP is a change in the synaptic strength based on the ordering of preand post-synaptic spikes (Bell et al., 1997). This mechanism has been proposed to explain the reinforce- Corresponding author. Tel.: +46 46 222 8543; fax: +46 46 222 4623. 1 Tel.: +33 4 7676 5625; fax: +33 4 7676 5619. E-mail addresses: tatyana@maths.lth.se (T.S. Turova), Alessandro.Villa@ujf-grenoble.fr, avilla@neuroheuristic.org (A.E.P. Villa). URL: http://www.neuroheuristic.org/. ment of synapses repeatedly activated shortly before the occurrence of a post-synaptic spike (potentiation) and the weakening of synapses strength whenever the pre-synaptic cell is repeatedly activated shortly after the occurrence of a post-synaptic spike (depression). There is evidence that the synaptic weights may follow a discrete distribution, in particular with respect to the glutamatergic NMDA-mediated receptors excitatory synapses (Montgomery and Madison, 2004). This mechanism might also drive selective synaptic pruning that lead to shape the newborn densely interconnected neural networks with removal of a large proportion of the initial connections (Huttenlocher, 1979; Huttenlocher et al., 1982; Bourgeois and Rakic, 1993). What is the role of the firing patterns presented at the begin of such process? What kind of topology can be obtained when the pruning process has stabilized? These are just few among many questions that can be raised in this topic (Song and 0303-2647/$ see front matter 2006 Elsevier Ireland Ltd. All rights reserved. doi:10.1016/j.biosystems.2006.05.019

T.S. Turova / BioSystems 89 (2007) 280 286 281 Abbott, 2001). Experimental investigations in animal models are limited to specific aspects of these questions with the techniques available nowadays but cannot address the global issues. The complexity of this problem has led the investigations to focus mainly on computational results of large scale neural networks modeling (Eriksson et al., 2003; Izhikevich et al., 2004; Iglesias et al., 2005a). This paper presents an original mathematical framework based on graph theory which is a first attempt to investigate this problem from a theoretical viewpoint. In order to keep this problem mathematically tractable it has been necessary to oversimplify a number of assumptions but the richness of the main questions was preserved. We present several results that can be compared with results obtained from computational simulations. In particular, we show that the phase transitions (along different parameters) take place in this type of network and we make first steps in the description of the phase diagram. Our results should be useful for further simulation analyses. 2. Network topology 2.1. Spatial distribution The network is a 2D lattice where each unit is assumed to be at the center of a relative map with coordinates x = 0, y = 0. Two types of units, excitatory and inhibitory, are laid down according to the following spatial distribution. The excitatory units are located at every point (vertex) of the graph Λ + = [ N, N] 2. The inhibitory units are more sparse because they are placed in the nodes of the enlarged, by factor 2, and shifted lattice Λ = (1/2; 1/2) + 2Λ +, thus Λ ={(1/2; 1/2) + (x, y) :x = 2k and y = 2n}. For large values of N the proportion of inhibitory units relative to the number of excitatory units tends to be 1:4. Then, the entire network is represented by the graph Λ = Λ + Λ. 2.2. Connectivity For any pair of units belonging to the network Λ, there is a synapse (u, v) between u Λ, the pre-synaptic neuron, and v Λ, the post-synaptic neuron, characterized by its synaptic weight. The set of the network connections at time t is denoted L(t). In other words let us define a (random) graph G(t) = (Λ,L(t)) with a set of vertices Λ and a set of directed edges L(t)at time t 0. For the initial conditions, at t = 0, let G(0) be a random graph on Λ such that any edge (u, v) is presented with a probability { q, if u v d, p 0 (u, v) = c q, otherwise, independent of other edges. Here c, q, and d are parameters of the model. We assume that c is much smaller than q, i.e., the units establish a dense net of short-range connections and a sparse net of long-range connections. Then L(0) is a random set of edges chosen with the above probabilities. 2.3. Synaptic weights At time t a synapse between u, the pre-synaptic neuron, and v, the post-synaptic neuron, is characterized by its synaptic weight which we denote w t (u, v). We set w t (u, v) 0, and denote for all t 0, if (u, v) / L(0), W(t) ={w t (u, v) :(u, v) L(0),t 0} the set of synaptic weights at time t that depend on the types of the pre- and post-synaptic units. The strength of the excitatory connections can take only four discrete states: zero (no connection), weak (α 0 ), medium (α 1 )or strong (α 2 ). Then, the set of excitatory connections at time t is described by w t (u, v) {0,α 0,α 1,α 2 }, u Λ +,v Λ, where 0 <α 0 <α 1 <α 2 1. The strength of the inhibitory connections can only take two values: w t (u, v) {0, γ}, u Λ,v Λ, where 0 <γ 1. For the initial conditions, at t = 0, it is assumed that any excitatory connection in W(0) = {w 0 (u, v) :(u, v) L(0)} has a medium strength, i.e.: { α1, if u Λ +, w 0 (u, v) = γ, if u Λ. Notice that the post-synaptic potentials may assume values in the range [0,M], corresponding to the synaptic strengths w t (u, v) multiplied by some positive constant M.

282 T.S. Turova / BioSystems 89 (2007) 280 286 3. Neuromimetic modeling The units of the graph correspond to neuromimes (oversimplified neuronal models) whose membrane potential X v (t) varies in the subthreshold range [0, 1]. A unit v fires a spike at time t if X v (t) = 1. Thus, the state of all network units at time t is defined by X(t) ={X v (t) [0, 1] : v Λ,t 0} with the initial state X v (0) = X v for all v, where X v are independent random variables uniformly distributed on [0, 1]. Immediately after firing, the state of a unit v is reset to a random state X v (t+), which is an independent copy of a random variable X v. Notice that the chance that X v (t+) = 1 is zero, thus there is always some positive period τ>0 between two consecutive spikes of one unit. Hence, we may consider that a refractory period here is at least τ>0. 4. Network dynamics The dynamics of the network is described by the Markov process (X(t),G(t),W(t)), t 0. 4.1. Decay of the excitatory synaptic weights To make our model mathematically tractable it is assumed that in the absence of any input the states of the neurons are kept constant. However, the weights of the excitatory connections may vary in time between states weak (α 0 ), medium (α 1 ) and strong (α 2 ). In the absence of any activity of the pre- and post-synaptic units it is assumed that the excitatory synapses tend to disappear, i.e., it is assumed that w t (u, v) is a random process with decreasing trajectories: from state α i it jumps to state α i 1 with intensity μ i, for i = 2, 1. This means that (in the absence of any activity) the synaptic weight remains at state α i for a random time, which is distributed exponentially with mean 1/μ i. In presence of activity the synaptic strengths can decrease or increase following a spike timing dependent plasticity (STDP) rule described later and vary between states α 0 and α 2. State zero is absorbing because the process w t (u, v) jumps from state α 0 to 0 with intensity μ 0 but once a synapse reached level 0 it cannot recover any other state and the connection is lost. Hence we define L(t) ={(u, v) L(0) : w t (u, v) 0} L(0), with the probabilities of edges as in the initial graph G(0). 4.2. Evoked activity Let us consider a time sequence T ={T 1,T 2,...}. At each time t T a set of units A 0 Λ receives a suprathreshold input such that those units generate a spike. We assume, that T n+1 T n τ. Note that τ should not be too large compare to 1/μ 0 + 1/μ 1 + 1/μ 2, since otherwise during the period without any activity the system can get with a high probability to the state where all the synaptic connections are zeros. Thus, let us set A 0 ={v Λ : X v (t) = 1}. This evoked activity is propagated to the target units along a path that follows the edges of the (random) graph G(t) = (Λ,L(t)). Let us examine some simple cases. 4.2.1. Firing of 1 inhibitory unit In this case A 0 contains only one vertex u Λ, i.e., only one inhibitory neuron is at state 1. For all v Λ such that there is an edge (u, v) in the graph G(t) and w(u, v) 0 the states of the units are updated and we set X v (t+) = max{0,x v (t) + w t (u, v)}. Thus, the inhibitory impulse may affect only the units that are receiving a direct input from the active unit u in graph G(t), and cannot propagate further in the network. 4.2.2. Firing of 1 excitatory unit In this case A 0 contains only one vertex u Λ +, i.e., only one excitatory neuron is at state 1. For all v Λ such that there is an edge (u, v) in the graph G(t) and w(u, v) 0 the states of the units are updated and we set X v (t+) = min{1,x v (t) + w t (u, v)}. Let us denote A 1 the set of units which fire due to the input from unit u: A 1 ={v Λ : X v (t) + w t (u, v) 1}\A 0. (1) The algorithm stops if set A 1 =. Otherwise, the algorithm continues recursively assuming the existence of sets A 0,A 1,...,A n for some n 1. Let us denote A n+1 the set formed by all the units which fire due to the impulses generated by the units with indices in n i=0 A i: A n+1 = v Λ : X v(t) + w t (v,v) 1 \ v n i=0 A i n i=0 A i. (2)

T.S. Turova / BioSystems 89 (2007) 280 286 283 The algorithm continues until minimal k such that A k+1 =. With this k it is possible to define the set A(t) = k i=0 A i. (3) This means that at time t+ all the units have a state defined by X v, if v A(t), { } X v (t+) = max 0,X v (t) + w t (v,v), if v/ A(t). v A(t+) Notice that as stated here the propagation of the impulses occurs instantaneously throughout the network and all units with indices in A(t) fire simultaneously at time t. However, the iterations may also be viewed as discrete steps at some micro-scale of time, where the time steps correspond to the indices 0, 1,...,k. 5. Results Our aim is to study the dynamics of the connectivity, i.e., the qualitative and quantitative features of the modifications of the graph G(t) = (Λ,L(t)) due to the STDP rule and to specific input patterns of activity. Firstly we derive a graph induced by the strong excitatory connections (i.e., synaptic weight α 2 ) after the first application of the STDP rule, and describe the architecture of this graph. Then we provide a (partial) description of the phase diagram on the space of parameters (c, q, α 1,γ), which will describe the size of the graph induced on G(0) by the STDP rule. At this stage of the study we do not present computational results which are currently under investigation. 5.1. Choice of the initial random graph G(0): parameters q and c 4.2.3. Firing of a set of units In general, if at time t only the units of some set A 0 have been activated, it is possible to proceed recursively as in the previous case, by replacing Eq. (1) with the following equation: A 1 = v Λ : X v(t) + w t (v,v) 1 \ A 0. (4) v A 0 4.3. Spike timing dependent plasticity rule Spike timing dependent plasticity (STDP) is a change in the synaptic strength based on the ordering of pre- and postsynaptic spikes. At this stage we present a very rough STDP rule that will be improved in future extensions of this work. Following the previous notation we can state that if v A i and v A j for some i<jit means that a post-synaptic unit v generates a spike immediately after the occurrence of a pre-synaptic spike of unit v.in this case it is assumed that the strength of the synapse is increased to the strong level, α 2. Conversely, if there are no post-synaptic spikes generated after a pre-synaptic spike the strength of the synapse is decreased to the weak level, α 0. In the current model the inhibitory connections are not modified by the STDP rule. Hence, the set of the synaptic weights is dynamically defined by the following equation: α 2, if v A i,v A j for i<jand w t (v, v ) > 0, w t+ (v, v ) = α 0, only if v A(t+) and w t (v, v ) > 0, w t (v, v ), otherwise. Let us remind that G(t) G(0) for all t, following the construction rule presented above (see Section 2.2). The graph G(0) is sparse if, with a high probability, any connected component has a size bounded uniformly in N. Then, even after the changes due to STDP the graph will be formed by bounded connected parts of the network. In order to study a dynamical phenomenon that propagates through a large part of the network, the parameters q and c should not be too small. Let us recall the basic facts from the percolation theory and the theory of random graphs which one should take into account here. If c = 0, then there is a constant 0 <q 0 = q 0 (d) < 1 such that if q>q 0 then, with a positive probability (independent of N), there is a connected path of short-range connections in G(0) from the origin (or any other unit) to the boundary of Λ.Ifq<q 0 then, with a probability one, any connected path of short-range connections in G(0) is uniformly bounded in N. The exact value of q 0 (d) is known only for d = 1: q 0 (1) = 1/2. Let q = 0. According to the theory of random graphs (Bollobás, 1985),ifc = c 1 /(2N) 2 with c 1 > 1 then G(0) has, with a high probability, a connected component which spans a positive fraction of all the vertices in Λ + (even when N ). If c = c 1 /(2N) 2 with c 1 < 1 then with a high probability, the largest connected component in Λ + is at most of order log N when N. Hence, if q<q 0 and c<1/(2n) 2 then, no matter what are the (5)

284 T.S. Turova / BioSystems 89 (2007) 280 286 other parameters, the induced network after the STDP rule will consist of only a very small fraction of units (exponentially small compared to the initial size). In the following results we shall consider parameters q and c outside of this area. 5.2. Propagation of a firing pattern A firing pattern is defined by a fixed set A 0 firing spikes at time T. The strength of the inhibitory connections is not modified by any STDP rule here, and for sake of simplicity we assume that at time T all the synaptic weights still have a medium strength, i.e., for any (u, v) L(0), w T (u, v) = w 0 (u, v) {α 1, γ}, as in the initial state. In this case Eq. (2) becomes A n+1 ={v Λ : X v (t) + α 1 {v n i=0 A i Λ + : (v,v) L(0)} γ {v n i=0 A i Λ : (v,v) L(0)} 1}\ n i=0 A i, where we simply have to calculate the number of excitatory and the inhibitory impulses for each unit, since the strengths of the connections are equal. (Note that here and elsewhere for any set A we denote A the number of elements in A.) Let us define now a random set A of units activated by A 0 according to Eq. (3): A = A(T ) = k i=0 A i, (6) where k is the minimal value for which defined above A k+1 =. Denote A + = A Λ + and A + i = A i Λ +, 0 i k, the subsets of the excitatory units in A. The sets A + i for all 0 i<kare not empty because the firing pattern can be propagated only through the excitatory connections. The set A + k can be empty, because A k may consist of only the inhibitory neurons. Notice that in the present scenario of propagation of firing from one single source, the (directed) cycles of edges cannot be found in the final structure. Cycles can be constructed, for example, in the course of an alternating activation of (at least) two different subsets of units. 5.3. Graph of strongly interconnected units Let us define G + the graph on the set A +, after application of the STDP rule (Eq. (5)) between its vertices, induced only by the strong connections, i.e., by the connections (u, v) for which w T + = α 2. Consider the set A + as an abstract set of vertices, not taking into account their coordinates in Λ. The sets A + 0,...,A+ k may be viewed as k + 1 parallel layers. Each unit may be viewed as the target of converging inputs, named here the in-degree, from the previous layer and the source of diverging outputs, named here the outdegree, projecting to the next layer. By construction the graph G + has the following basic properties: there is at least one path from the initial set A + 0 and at least one edge from the set A + i 1 to any vertex of any set A + i, i 1; the edges may go from the vertices of a set A + i to the vertices of a set A + j if and only if i<j; there are no edges between vertices of the same set A + i for any i. Let us take into account the metric properties of the original space Λ and distinguish the two following subgraphs of G + : G + S is the subgraph that contains all the short-range edges in A +, including their endpoints, and V S is its set of vertices. G + L is the subgraph that contains only long-range edges in A +, including their endpoints, and V L is its set of vertices. Notice that V S and V L may have common vertices. The parameters of the model determine very different structures and relations between subgraphs G + S and G+ L. 5.4. Effect of connection strengths α 1 and γ We refer here to the propagation of a firing pattern as defined in Section 5.2. We assume that A 0 is small. 5.4.1. Strong excitatory connections: α 1 1 and γ = 0 In the marginal case, if α 1 = 1 and γ = 0, the graph G + is the subgraph of G(0): it consists of A 0 and all the units of Λ + (and the edges between them) which are connected to A 0 in G(0). If γ = 0 the probability that a unit will fire due to a firing of a single neighbour is qα 1, if a short-range neighbour fires, and it is cα 1 if a long-range neighbour fires. This analysis can be extended to consider the probability of firing due to the simultaneous firing of several units, which leads to two different phases: if parameters (α 1,γ) are in the vicinity of point (1, 0) the structure of G + is basically the same as the subgraph of G(0) connected to A 0 ; if 0 α 1 α 0 1 (for some 0 <α0 1 < 1) then, no matter what are the values of other parameters, the graph

T.S. Turova / BioSystems 89 (2007) 280 286 285 G + will never span a positive fraction of the entire network. 5.4.2. Weak excitatory connections: α 1 0 Let us assume that the synaptic weight α 1 is weak. In this case a unit must receive enough inputs to reach the threshold of discharge in order to generate a spike, i.e., a unit must have a sufficiently high in-degree. Denote ν in the minimal in-degree which yields, with a high probability, the firing of a unit. Obviously, ν in is increasing when α 1 is decreasing. By the assumptions on the probabilities p 0 (u, v), we have c q, where c is the probability of a long- and q is the probability of a short-range connection (see Section 2.2). This implies that the main contribution into ν in will be from the short-range connections. Furthermore, for small values of α 1 the subgraph G + L will not be generated at all. The graph G + can be further described as follows: the only vertices with in-degree zero can be found in A 0 ; there is a high probability that the in-degree of any other vertex is at least ν in. Hence, for weak synaptic weights α 1, the first firing pattern propagates through a forest-like (converging) graph with the maximal length of the connected path of order log A 0, which is also the order of value k. 5.4.3. Critical excitatory connections Based on our preliminary analysis above we can derive that for any fixed values c 1 > 1, q>q 0 and γ = 0 there is a critical value 0 <α cr 1 (0) < 1, such that if α 1 is above this value then the graph G + spans a positive fraction of all the units in the network. We conjecture, that for all 0 <γ<1 there is a critical value α cr 1 (0) α cr 1 (γ) 1, above which G+ spans a positive fraction of all the units in the network. It is plausible to think that α cr 1 (γ) is an increasing function of γ. It is an interesting open question whether the strict inequality α cr 1 (γ) < 1 holds for all 0 <γ<1. In other words, whether indeed for any γ there is a parameter which yields a positive fraction of the connected units in the network. 5.5. Effect of the initial firing pattern A 0 The interesting result here is that the geometry of the set A 0 affects directly only the subgraph G + S. Let the number of the excitatory vertices A + 0 =m N2 be fixed. Let us consider two marginal examples of different configurations of set A + 0 in Λ. The distance between any two vertices in A + 0 is greater than d. Thus, A + 0 most likely does not generate a subgraph G + S at all. The set A + 0 covers a circle around some vertex. Thus, for large values of m the vertices in A + 0 at distance greater than d from the boundary of the set A + 0 do not contribute to the formation of G + S. In both cases the subgraph G + L remains the same because it depends only on the size m of A + 0. 6. Discussion This paper has presented an original mathematical framework to study the dynamics of a model of neural networks with embedded spike timing dependent plasticity. The presynaptic signals that lead to a spike enhance the strength of an excitatory synapse by STDP, thereby reinforcing a specific path of the connectivity graph. Recent study suggest that some inhibitory synapses may be modified in a spike timing dependent manner, somewhat similar to what is observed for the excitatory (Harkany et al., 2004). In the case of inhibitory STDP, the increase in synaptic strength would lead to an effect opposite to the storage or reinforcement of the activity patterns that caused it. We do not rule out that the inhibitory STDP may play an important role in shaping preferred paths through the graph, but at this stage of our model it is premature to introduce this additional parameter. We have emphasized the importance of the initial parameters, in particular of the initial topology that describes patterns of short- and long-range excitatory connections. This is in agreement with well known data of the cerebral cortex anatomy (Szentágothai, 1975; Braitenberg and Schüz, 1988). The cortico-cortical connections are characterized by short-range dense columnar and areal projections and long-range sparse interareal and interhemispheric connections (Greilich, 1984). In the current framework, with relatively weak excitatory connections, our results show that the degree of the diverging projections (out-degree) is smaller than the degree of converging inputs (in-degree) for both graphs of short- (G + S ) and long-range (G+ L ) connections. Indeed we should consider that the smaller the medium synaptic weight, α 1, the larger should be the in-degree and smaller the out-degree of any vertex in A + \ V 0. This means that a unit may successfully contribute only to the firings of a small number of other units, but it requires many more inputs to become activated. This is in good agreement with the computational results obtained with similar models (Iglesias et al., 2005b).

286 T.S. Turova / BioSystems 89 (2007) 280 286 The initial firing pattern A 0 plays also an important role. To get more excitation in the network it is appropriate to have A + 0 composed of assemblies of units grouped in patches of some diameter d placed next to each other at certain distances e. This parameter e should be chosen such that the excitation may propagate from one group to the other throughout the entire network. We suggest that the diameter d of A + 0 plays a key role for the study of G(t) knowing the structure of graph G +. In conclusion, the study of neural networks dynamics with the help of graph theory may offer new perspectives for the identification of critical parameters of the models. Future simulations should consider these results to guide the build-up of more accurate models and focus the exploration of the parameter space on targeted areas. Acknowledgments The authors thank the referees for the helpful remarks. The research of T.T. was supported by the Swedish Natural Science Research Council and the Mathematical Science Research Institute, Berkeley, USA. References Bell, C.C., Han, V.Z., Sugawara, Y., Grant, K., 1997. Synaptic plasticity in a cerebellum-like structure depends on temporal order. Nature 387, 278 281. Bollobás, B., 1985. Random Graphs. Academic Press, London. Bourgeois, J., Rakic, P., 1993. Changes of synaptic density in the primary visual cortex of the macaque monkey from fetal to adult stage. J. Neurosci. 13, 2801 2820. Braitenberg, V., Schüz, A., 1988. Cortex: Statistics and Geometry of Neuronal Connectivity, 2nd ed. Springer Verlag, Berlin. Eriksson, J., Torres, O., Mitchell, A., Tucker, G., Lindsay, K., Rosenberg, J., Moreno, J.M., Villa, A.E.P., 2003. Spiking neural networks for reconfigurable poetic tissue. Lecture Notes Comput. Sci. 2606, 165 174. Greilich, H., 1984. Quantitative analyse der cortico-corticalen fernverbindungen bei der maus. PhD Thesis, University of Tübingen. Harkany, T., Holmgren, C., Hartig, W., Qureshi, T., Chaudhry, F.A., Storm-Mathisen, J., Dobszay, M.B., Berghuis, P., Schulte, G., Sousa, K.M., Fremeau, R.T.J., Edwards, R.H., Mackie, K., Ernfors, P., Zilberter, Y., 2004. Endocannabinoid-independent retrograde signaling at inhibitory synapses in layer 2/3 of neocortex: involvement of vesicular glutamate transporter 3. J. Neurosci. 24, 4978 4988. Huttenlocher, P.R., 1979. Synaptic density in human frontal cortex developmental changes and effects of aging. Brain Res. 163, 195 205. Huttenlocher, P.R., de Courten, C., Garey, L.J., Van der Loos, H., 1982. Synaptogenesis in human visual cortex evidence for synapse elimination during normal development. Neurosci. Lett. 33, 247 252. Iglesias, J., Eriksson, J., Grize, F., Tomassini, M., Villa, A.E.P., 2005. Dynamics of pruning in simulated large-scale spiking neural networks. BioSystems 79, 11 20. Iglesias, J., Eriksson, J., Pardo, B., Tomassini, M., Villa, A.E.P., 2005. Emergence of oriented cell assemblies associated with spiketiming-dependent plasticity. Lecture Notes Comput. Sci. 3704, 59 68. Izhikevich, E.M., Gally, J.A., Edelman, G.M., 2004. Spike-timing dynamics of neuronal groups. Cerebral Cortex 14, 933 944. Montgomery, J., Madison, D., 2004. Discrete synaptic states define a major mechanism of synapse plasticity. Trends Neurosci. 27, 744 750. Song, S., Abbott, L.F., 2001. Cortical development and remapping through spike timing-dependent plasticity. Neuron 32, 339 350. Szentágothai, J., 1975. The module-concept in cerebral cortex architecture. Brain Res. 95, 475 496.