Excitatory, inhibitory and structural plasticity produce correlated connectivity in random

networks trained to solve paired-stimulus tasks

Running head: Structural correlations from learning

Paul Miller1,2,3 and Mark Bourjaily2,3

1
Department of Biology, 2 Neuroscience Program and
3
Volen National Center for Complex Systems, Brandeis University, Waltham, MA 02454

Keywords: structural plasticity; connectivity; Hebbian learning; network; simulation; correlations;

STDP; inhibitory plasticity.

Address correspondence to:

Dr. Paul Miller

MS013, Volen National Center for Complex Systems,

Brandeis University, 415 South Street,

Waltham, MA 02454

USA

1
 Abstract

Many cognitive tasks require animals to produce a distinct behavioral response to specific pairs of

stimuli, different from responses to the constituent individual stimuli. In our analyses of trained

networks of spiking neurons, we found those capable of producing reliably correct behavioral responses

contained significant numbers of neurons that were supra-linearly responsive to specific pairs of stimuli.

In this work we investigate how the underlying structural features of trained networks correlate with task

performance. Thus, we model an initially randomly connected recurrent network of spiking neurons that

receive random inputs, in order to investigate what synaptic plasticity mechanisms sculpt the network

into a functional one with high stimulus-pair selective neural activity. In order to model a pair-

associative task, biconditional discrimination, we stimulate the network via consecutive paired-stimulus

combinations (A+B, C+D, A+D and C+B) of four inputs (A, B, C and D). We examine the effect of a

Hebbian form of spike-timing-dependent plasticity (STDP) between excitatory cells, as well as long-

term potentiation of inhibition (LTPi). LTPi directly modifies inhibitory synapses onto excitatory cells,

which generates the strong stimulus-pair selectivity necessary to solve the task. Without directly

modifying recurrent excitatory synapses, LTPi can alter correlations between excitatory neurons. When

we include an implementation of structural plasticity – where connections between uncorrelated

excitatory cells are removed and replaced by new random connections during training – the correlations

in structural connectivity between excitatory cells observed in vitro arise most often in networks with

the inhibitory plasticity mechanism, LTPi. Moreover, we find that only networks able to extract

stimulus associations – a building block of computation and essential role for cortex – produce an

excessive clustering of synaptic connections, as observed in cortical slices.

2
Introduction

Connections between neurons are not randomly distributed, but contain correlations (Song et al.,

2005; Lefort et al., 2009). In particular, a study of connections among cells in a small region of visual

cortex demonstrated that bidirectional connections between pairs of neurons are much greater than

expected by chance, given the measured probability of individual connections (Song et al., 2005). The

result does not simply arise from differing distances between cells (nearby cells have greater

connections probability) because all of the cells measured had overlapping dendritic and axonal arbors.

Moreover, when the authors analyzed triplets of cells, they found a number of three-cell connection

patterns or “motifs” to be more abundant than expected by chance, even after accounting for the excess

of bidirectional connections (Song et al., 2005).

The architecture of connections between neurons is not static, but can vary on a timescale of

hours (Minerbi et al., 2009) or days (Trachtenberg et al., 2002; Holtmaat et al., 2005), because of

ongoing formation and loss of dendritic spines and axonal contacts onto those spines (Yuste and

Bonhoeffer, 2004). While the determinants of spine retention versus spine loss have not been

characterized as well as the determinants of changes in synaptic strength, recent evidence suggests the

requirements are similar (Toni et al., 1999; Alvarez and Sabatini, 2007; Becker et al., 2008; Wilbrecht et

al., 2010) and synaptic strength correlates with synaptic size, while it is the smaller spines that are most

likely to disappear (Holtmaat et al., 2006; Becker et al., 2008). Moreover, the correlations observed in

the binary existence of synapses between cells matches the correlations observed in the strengths of

synapses between cells (Song et al., 2005). Recent modeling work by others, (Clopath et al., 2010) has

shown that a Hebbian-like plasticity mechanism, voltage-dependent STDP, produces the observed

correlations in synaptic strength – if a connection from cell i to j is strong, then the reverse connection,

from j to i is more likely than average to be strong than weak. If one assumes that weak connections

3
disappear, then the natural effect is an excess of bidirectional connections between cells, compared to

that expected by chance.

In this paper, using numerical simulation, we study how a network, initially containing sparse

and random recurrent connections among spiking neurons, can be trained to produce the observed

structural correlations. By studying a number of different types of network and plasticity mechanisms,

we find the experimentally observed correlations can arise so long as: (1) inputs are correlated; (2)

synaptic plasticity within the network is appropriate to produce cells responsive to the input correlations;

and (3) a form of Hebbian-like structural plasticity is present.

The particular training protocol we used is one that requires the network to produce an

Exclusive-Or (XOR) logical response to pairs of activated stimuli (Sakai and Miyashita, 1991; Dusek

and Eichenbaum, 1998; Harris et al., 2008; Harris et al., 2009). Production of such responses is non-

trivial, since a correct set of responses cannot be achieved by performing a selection or choice according

to a linear combination of the inputs. Networks capable of XOR-logical responses are of interest,

because by combining such networks, one can in principle produce a solution to any computable

problem. We find that the manner in which the network achieves XOR-logic is by producing cells

responsive to particular pairs of stimuli. Such selectivity to particular combinations of stimuli is a well-

known function of sensory areas (Desimone et al., 1984; Ito et al., 1995; Baker et al., 2002), which,

when coupled with a method for producing invariance, provides a framework for general feature and

item detection (Serre et al., 2007). Thus, we believe the task and ensuing network response simulated in

our model networks is representative of many of the computations carried out by circuits of neurons

across many areas of the brain.

Materials and Methods

4
The overall network structure and training protocols are described in a prior publication (Bourjaily,

2010). In this paper we add structural plasticity to the networks studied previously and analyze the

resulting changes in network structure in a manner similar to that used by Song et al. (Song et al., 2005)

for the connections within cortical slice data.

Network Inputs

Pairs of stimuli activated together (either A+B, C+D, A+D, C+B) produced inputs as Poisson

spike trains. The number of independent Poisson spike trains per stimulus was 2 in our default network,

but varied in comparisons across networks (2, 4, 6, 10 or 20) according to the label “Number of

Independent Input Groups per Stimulus.” Rate of the Poisson spike trains was 480Hz divided by number

of input groups. By increasing the number of input groups, we reduced the correlations in connectivity

from afferent cells, so for example with 20 input groups any stimulus produced 20 Poisson spike trains

of 24Hz, with independent sets of connections to cells within the associative layer. Whereas, with 2

input groups, the 2 trains of 240Hz Poisson spikes are equivalent to 20 Poisson spike trains of 24Hz, but

the 20 trains grouped into 2 sets of connections, with connections 100% correlated within a set. In this

paper we do not address structural plasticity of the input connections, which we expect to determine the

actual level of input correlations.

Each cell in the associative network was assigned any individual input spike train with a

probability of 1/5 in the default network, though the probability varied (1/2, 1/3, 1/5, 1/10 or 1/20) when

assessing multiple networks.

In a subset of examples (Figures 8 and 9A,C) we train the network, not with a task involving

paired inputs, but by presenting a single stimulus at a time. Total time of inputs is unchanged, but neural

responses are sparser, producing greater selectivity (fewer cells active at any one time and any cell is

active to fewer inputs).

Neuron Properties

5
 We use leaky integrate-and-fire (LIF) neurons (Tuckwell, 1988) defined by the leak

conductance, gL, synaptic conductances gAMPA, gNMDA, gGABA, and a refractory conductance, gref, resting

potential (i.e. leak potential) VL and threshold potential, Vth. The threshold potential is dynamic: it

increases to a maximal value of 150mV immediately after a spike and decreases exponentially (with

time constant 2ms) to a cell-dependent base value (given below) between spikes. The membrane

potential, V, evolves according to the equation:

dV
C = gL (VL − V ) + gGABA (t)(VI − V ) + gNMDA (t)(VE − V ) FNMDA (V ) + gref (t)(Vref − V ) + gext
E I
(VE − V ) + gext (VI − V )
dt

VE and VI are reversal potentials for excitatory and inhibitory currents respectively. Rather than a hard
€
reset, following a spike, we mimic delayed rectifier potassium currents with reversal potential, Vref = -

70mV, via a refractory conductance, gref, which increases immediately following a spike by δgref =

150µS and decays to zero exponentially with a cell-dependent time constant, τref.

Noise:

We model noise as independent excitatory and inhibitory synaptic conductance variables, drawn from a

uniform distribution [0 1.2mV/s1/2 ].

Associative layer parameters

LIF neurons had a mean leak reversal potential of VL = -70mV +/- 2.5mV, membrane time

constant of τm = 10ms +/- 0.75ms and leak conductance of gL = 35µS +/- 1µS. Excitatory neurons had a

firing threshold of Vth = -50mV +/- 2mV, a reset voltage of Vreset = -60mV +/- 2mV, and a refractory

time constant of τreset = 2ms +/- .25ms. Inhibitory neurons had a firing threshold of Vth = -50mV +/-

2mV, a reset voltage of Vref = -60mV +/- 2mV, and a refractory time constant of τreset = 1ms +/- .25ms.

Heterogeneity of these parameters was drawn from uniform distributions with the given ranges.

Decision layer parameters (Wang, 2002)

6
 Excitatory LIF neurons had a mean leak reversal potential of VL = -70mV, membrane time

constant of τm = 20ms, and leak conductance of gL = 35µS. Excitatory neurons had a firing threshold of

Vth = -48mV, a reset voltage of Vreset = -55mV, and a refractory time constant of τref = 2ms. Inhibitory

LIF neurons had a mean leak reversal potential of VL = -70mV, membrane time constant of τm = 10ms,

and leak conductance of gL = 30µS. Inhibitory neurons had a firing threshold of Vth = -50mV, a reset

voltage of Vreset = -55mV, and a refractory time constant of τref = 1ms.

Synaptic interactions

Synaptic currents were modeled by instantaneous steps after a spike followed by an exponential

decay described by the equation below (Dayan and Abbott, 2001).

Recurrent excitatory currents with reversal potentials, VE = 0mV, were modeled as mediated by AMPA

receptors (τAMPA = 2ms) and NMDA receptors (τNMDA = 100ms). Inhibitory currents with reversal

potential, VI = -70mV, were modeled as mediated by GABAA receptors (τGABA = 10ms). Conductance of

NMDA receptors was modified by the voltage term FNMDA(V) (Jahr and Stevens, 1990):

FNMDA (V ) = 1 {1+ [ Mgext

2+
] exp(−0.062V ) /0.00357} with the external magnesium concentration, [Mgext2+]
= 1mM.
€ Synaptic Input Sparseness and Correlations

In order to investigate the robustness of each learning rule, we examined their effects on sets of

25 different networks with each set explored across six network regimes. We examined how the

sparseness and correlations of input groups affected both the initial selectivity of a network and how the

network responds to each of the synaptic plasticity rules. Input sparseness is defined via the probability

of any input group projecting to any given cell. As input connection probability increases, sparseness

7
decreases. We used the following five values for input connection probability: 1/2, 1/3, 1/5, 1/10 and

1/20.

We produced different degrees of input correlations by altering the number of independently

connected input groups of cells per stimulus, using 2, 4, 6, 10 or 20 independent groups. Each input

group produced independent Poisson spike trains with a mean firing rate defined by:

= 480Hz/Number of Input groups (e.g. 10 input groups of 48Hz). Correlations weakened

progressively as the number of inputs increased due to the increasing number of independent input

Poisson spike trains producing the same overall spike rate.

Five levels of input sparseness, combined with five different degrees of input correlations led to

25 variant networks in each regime.

Our initial network possessed no structure in its afferent connections and in its internal recurrent

connections. Random connectivity produced cell-to-cell variability since no two cells receive identical

inputs. Such heterogeneity of the inputs across cells led to a network of neurons with diverse stimulus

responses.

Associative layer connectivity

Excitatory-to-excitatory connections are sparse with 10% connection probability. Initially

connections are random and uncorrelated. Inhibition is feedforward only, so there are no excitatory-to-

inhibitory connections. Inhibitory-to-Inhibitory connections are all-to-all. Finally, Inhibitory-to-

excitatory synapses connect randomly with a probability of 25%. Initial excitatory to excitatory

synaptic strength is taken from a uniform distribution, with a mean value of W0=0.05 and range of +/-

50% about the mean. These simulations were carried out in a network with 320 excitatory and 80

inhibitory neurons. Connections to the decision layer are initially all-to-all from excitatory neurons with

a uniform strength of DW0=0.075.

Decision layer connectivity

8
 The decision-making network based on prior models (Wang, 2002) is composed of two

excitatory pools (each containing 200 cells) and two inhibitory pools (each containing 50 cells.

Excitatory to excitatory synaptic strength is W0=0.25. Connections within each pool are all-to-all.

Cross-inhibition is direct from each inhibitory pool to the opposing excitatory pool, which generates

winner-take-all activity so that only one pool is stable in the up state (active). Network bistability is

generated by strong inhibition and self-excitation.

The decision-making network receives a linear ramping input initiating at the start of the cue and

continues until the end of the cue where it reaches its maximal value of gurgency=5µS at the end of the

cue. This input is adapted after the “urgency-gating” model (Cisek et al., 2009), and it ensures that a

decision is made each trial.

Plasticity rules

For all connections, changes in synaptic strength are limited to a maximum of 50% per trial, while

across all trials, synaptic strength is bounded between zero and 20 times W0, the initial mean synaptic

strength.

Structural Plasticity

Our rule for structural plasticity assumes that synapses disappear if there is little Hebbian-like causal

correlation between presynaptic and postsynaptic spikes. The synapses are replaced at random. We

consider three types of random replacement that maintain the total number of connections at a constant

value:

(1) Select at random a new presynaptic cell, keeping the same postsynaptic cell.

(2) Keep the same presynaptic cell, but select at random a new postsynaptic cell.

(3) A 50% choice of process (1) or process (2).

The criterion for selection of synapses to be removed is based on three parameters:

9
 (1) The width of a temporal window for coincidence of spiking – if a postsynaptic cell spikes

within such a temporal window following a presynaptic spike, the synapse between the two

cells is more likely to be retained.

(2) The frequency of such coincidences necessary to prevent removal of the synapses.

(3) The number of trials without sufficient numbers of coincident spikes to result in removal of

the synapse.

The main effect of changing these parameters is to alter the number of synapses that get removed across

trials, rather than changing from removal of one set of synapses to another set. Thus, the magnitudes of

correlations are affected by parameter changes, but not the overall sign and directions of correlations.

The specific implementation of structural plasticity under our standard parameter set is to update

a parameter, Sij, for each synapse following each trial, and remove the synapse if Sij falls below a

threshold, T = -4. Any new synapse and all synapses at the beginning of the simulation are initialized

with Sij = 0. Within a trial, for each pair of presynaptic and postsynaptic spikes at times ti and tj

[ ]
respectively, with tj > ti, we increase Sij according to: Sij  Sij + exp −( t j − t i ) / τ Struc , where

τ Struc = 25 ms. The necessary rate of coincident spiking to prevent loss of a synapse is given by the value
€ trial: S  S − R . Thus a new synapse can be removed
R = 0.5, which is subtracted from Sij after each ij ij

€
after 8 trials (because T/(-R) = 8) if its presynaptic and postsynaptic neurons produce no causally

coincident spikes in these trials. The more €

coincident spikes in the past history of the synapse, the more

trials it can survive without coincident spiking, thus strong synapses are more stable. In our simulations,

since the total numbers of inputs are the same for every block of 4 trials, the history-dependence is not

an important factor (if we set an upper bound on Sij of Sijmax = 4R = 2, our results are identical).

LTPi

LTPi is modeled after (Maffei et al., 2006): LTPi occurs when an inhibitory cell’s fires, but the

excitatory cell is depolarized and silent. If the excitatory cell is co-active (i.e. spiking), then there is no

10
change in the synapse strength. We refer to this as a veto effect in our model of LTPi. Any excitatory

spike within a window of +/- 20ms for an inhibitory spike will result in a veto. For each inhibitory spike

(non-vetoed) the synapse is potentiatiated by idW=0.005.

LTPi was reported experimentally as a mechanism for increasing (but not decreasing) the strength of

inhibitory synapses in cortex (Maffei et al., 2006). To compensate for the inability of LTPi to depress

synapses, we use multiplicative post-synaptic scaling (Turrigiano et al., 1998a) for homeostasis at the

inhibitory-to-excitatory synapses. We explicitly model the post-synaptic depolarization required by

LTPi by defining a voltage threshold that the post-synaptic excitatory cell must be above in order for

potentiation to occur. We used a value of -65mV, which is 5mV above the leak reversal. Finally, we

include a hard upper bound of inhibitory synaptic strength, such that those cells most strongly inhibited

(so being less depolarized as well as not spiking) in practice receive no further potentiation of their

inhibitory synapses.

Triplet STDP

Triplet STDP was modeled after the rule published by Pfister & Gerstner 2006 (Pfister and

Gerstner, 2006). Their model includes triplet terms, so that recent postsynaptic spikes boost the amount

of potentiation during a “pre-before-post” pairing, while recent presynaptic spikes boost the amount of

depression during a “post-before-pre” pairing. Specifically when

> 0,

< 0,

We use the parameters cited from the full model “all-to-all” cortical parameter sets in the paper.

The amplitude terms are doublet LTP A2+=5*10-5, doublet LTD A2-=7*10-3, triplet LTP A3+=6.2*10-3,

and triplet LTD A3-=2.3*10-4. The time constants we used are τ2+=16.68ms, τ2-=33.7ms, τy=125ms, and

11
 τx=101ms. These parameters generated an LTD-to-LTP threshold for the post-synaptic cell of 20Hz,

above which uncorrelated Poisson spike trains produce potentiation and below which they produce

depression. For every spike that updates the synapse the synaptic strength changes by dW=0.005.

Homeostasis by multiplicative synaptic scaling

Synapse stability is maintained by multiplicative post-synaptic scaling (Turrigiano et al., 1998a;

Turrigiano and Nelson, 2000; Renart et al., 2003) with synaptic strengths updated on a trial-by-trial basis

as ΔW ij = εW ij ( rgoal − rj ) , where the mean rate of the postsynaptic cell, j, is rj and its goal rate is rgoal. We

use the parameters, ε=0.01 (inhibitory-to-excitatory synapses), ε=0.0001 for input to and recurrent
€ excitatory synapses. Goal rates for all types of cell and synapse were taken from a uniform distribution

with mean 8Hz and range +/- 50%.

Stimulus-pair selectivity metric:

Stimulus-pair selectivity, SPSi, defines for each excitatory neuron, i, its selectivity for one

stimulus-pair over the other three stimulus-pairs. SPSi is the maximum firing rate of neuron i, minus its

mean response across all stimuli, normalized by its mean response: S PS i = ri ( max
− ri ) ri . The

network’s stimulus-pair selectivity value, <SPS> is the mean of SPSi taken across all excitatory cells.

Control Network of Cell Assemblies €

As a control, we produced a network of 320 excitatory cells, randomly assigning each cell to one

of four assemblies. We then assigned synaptic connections randomly, but with connection probability

twice as high (0.2) within an assembly compared to between assemblies (0.1). Connection strengths

were selected randomly from a Gaussian distribution with mean of 1, standard deviation 0.5 for within

assembly connections, and with a mean of 2, standard deviation 0.5 for between assembly connections

(negative values were reselected). We did not simulate the activity of such a network, but assessed how

the correlations among connections produced in such a simple manner match the observed data.

Numerical Simulations

12
 Simulations were run for 400 trials, using the Euler-Maruyama method of numerical integration

with a time step, dt=.02ms. All simulations were run across at least four random instantiations of

network structure, cell and synapse heterogeneity, and background noise. Simulations were written in

C++ on Intel Xeon machines. Matlab R2010a was used for data analysis and visualization.

Results

Differences between types of structural plasticity

Our formulations of structural plasticity are Hebbian, requiring a presynaptic spike to precede a

postsynaptic spike sufficiently often to prevent synapse removal. Since we assume that cells have no

mechanism to detect correlations in their activity until a connection forms, we produced new

connections at random, via three different methods. In the first method, once a synapse is removed, we

keep track of the postsynaptic cell and provide it a new, randomly chosen afferent connection. This

method is analogous to keeping the number of dendritic spines constant but allowing spine mobility to

produce new connections. Thus, the number of incoming, afferent connections per cell remains constant

while the number of outgoing, efferent connections can vary greatly across the network (Figure 2A-B).

The second method is a mirror of the first, as we keep track of the presynaptic cell and produce a new

postsynaptic partner. This is analogous to maintaining the number of axonal boutons, but allowing them

to move and connect to new cells. Thus the number of efferent connections does not change, but a broad

distribution of numbers of afferent connections arises (Figure 2C-D). In the third method, neither

number of afferents nor efferents is fixed (Figure 2E-F) as with 50% probability we switch the

preynaptic cell and with 50% probability we switch the postsynaptic cell. All of these methods assume

processes arising from one cell move to form a new partner. The other logical possibility of choosing a

totally new random pair of cells, is less biological, and produces results equivalent to our third method

(data not shown).

13
 We note that following structural plasticity, some cells may lack efferents or afferents or both,

depending on implementation. Such a result is not unreasonable when considering the relatively small

number of excitatory cells (320) and the paucity of activated stimuli that are simulated.

All three implementations of structural plasticity produced similar changes in the numbers of

bidirectional connections, but produced differences in the patterns of connections – or motifs – among

triplets of cells (Figure 3A-F). The most obvious difference between switching presynaptic versus

switching postsynaptic cells in the structural plasticity rule appeared in the motifs numbered 4 and 5

(Figure 3G). If the presynaptic cell was switched, so that some cells had more efferent projections than

others then motif 4 arose more often, whereas if the postsynaptic cell was switched all cells had the

same number of efferent projects but some received more connections and motif 5 was more common.

In vitro data (Song et al., 2005) demonstrate an excess of motif 4. Our third implementation produced an

excess of motif 4, but also an excess of motif 11 – and all other connected triplets, motifs 10 and higher

– in agreement with slice data (and unlike the first two methods). Thus, the third implementation was

our default method for the rest of this work.

Structural plasticity can increase numbers of bidirectional connections and connected triplets

Initially, internal connections were random (with 10% probability), so numbers of bidirectional

connections or connected triplets of cells were at chance level, given the total number of connections.

Following 400 trials of presentation of paired stimuli in our default network with both long-term

potentiation of inhibition (LTPi) and triplet-STDP (Pfister and Gerstner, 2006), the ratio of number of

pairs with bidirectional connections to the expected number increased from to 2.3 ± .2 (4 independent

simulations). The numbers of fully connected triplets (motifs 10 or higher, Figure 3G) increased by the

same factor, 2.3 ± .2 relative to chance. Intriguingly, if, while retaining structural plasticity and triplet-

STDP, we did not include LTPi, so that inhibition to excitatory cells remained unchanged from their

14
initial value, then the ratios fell to 1.4 for both connected doublets and triplets. Moreover, the ratios for

networks with LTPi but no triplet-STDP were 1.8 for doublets and 1.9 for triplets, showing that

inhibitory plasticity allowed structural plasticity to sculpt excitatory connections more than a Hebbian

form of excitatory plasticity. In fact, with neither triplet-STDP nor LTPi, changes produced by structural

plasticity alone led to ratios of 1.2 for connected doublets and 1.3 for connected triplets – not much less

than the network with triplet-STDP included.

To assess the generality of our result, for each combination of plasticity mechanism, we

produced 25 types of network, differing in their input structure via 5 values of connection probability

per input (1/20 to 1/2) and 5 values for the number of different random sets of input connections

activated per stimulus (2 to 20). Figure 4 demonstrates that addition of LTPi during training with

structural plasticity (C compared to A; D compared to B) increased the number of pairs of neurons with

bidirectional connections. Moreover, in this type of protocol with paired stimuli, we found that triplet-

STDP often reduced the number of such pairs (B compared to A; D compared to C). These results,

showing an increase in ratio by LTPi and often a decrease in ratio by triplet-STDP, were reproduced

when analyzing the excess of fully connected triplets (Figure 5).

Clustered connectivity correlates with task performance

In a prior publication (Bourjaily, 2010) we showed the necessity of high stimulus-pair selectivity

among cells in the associative network (Figure 1B) to produce reliable behavior as determined by the

trained output of a decision-making network. Moreover, in many of these networks, LTPi was necessary

to produce high stimulus-pair selectivity. Thus, in our protocol, the reason LTPi leads to more clustered

connections among cells, may be because such clustering of connections arises when subsets of cells are

active strongly together for a small subset of stimuli – a feature of cell assemblies. Such behavior would

be indicated by a high stimulus-pair selectivity index for the network, and would correlate with reliable

decision-making and high accumulation of reward in the task.

15
 Indeed, Figure 4 demonstrates, across 100 different networks (four random examples of each of

25 types) the high correlation between either paired-stimulus selectivity (A, B) or accumulated reward

(C, D) and either the excess of bidirectional connections (A, C) or fully connected triplets of cells (B,

D).

Figures 4A, B suggest a threshold value of selectivity (approximately 1.2), above which the neural firing

is sufficiently structured to produce the high clustering of connections. Since reward accumulation also

relies on selective responses, we find all of these features are significantly correlated with each other as

follows: bi vs selectivity, ρ=.60, p=3x10-11; bi vs reward, ρ=.68, p=5x10-15; tri vs selectivity, ρ=.54,

p=5x10-9; tri vs reward, ρ=.71, p=1x10-16; selectivity vs reward, ρ=.47, p=6x10-7; bi vs tri, ρ=.98,

p=3x10-74.

Synaptic strengths

We assessed changes in synaptic strength of excitatory to excitatory connections, in our default

network undergoing both triplet-STDP and LTPi with structural plasticity. Figure 7A indicates the non-

Gaussian distribution of connection strengths, with a long tail to higher strengths. Some synapses not

visible on the graph were above 0.3 in strength. The hard bound of 0.5 (10 times initial strength) was not

reached by any synapses in these simulations. Such a non-Gaussian distribution is observed

experimentally (Song et al., 2005; Lefort et al., 2009). In our simulations, low-strength synapses did not

arise, as those synapses weakened by triplet-STDP were also those selected for elimination by our

implementation of structural plasticity. New synapses were introduced with an initial strength of 0.05,

leading to the peak in distribution near a value of 0.05.

In vitro data shows that bidirectional synapses are stronger than unidirectional ones (Song et al.,

2005), a feature reproduced in a network model with a different protocol from ours (Clopath et al.,

2010). In some very sparse networks – those with input connection probabilities of 1/20 – highly

16
significant increases in bidirectional connection strengths were apparent: for example, the mean ratio of

bidirectional to unidirectional synaptic strength was 1.08 (p<10-100 ) in the sparsest networks with 4

input groups per stimulus. In our default network, we found no significant evidence for such an increase.

Figure 7B demonstrates, in our default network, a high degree of correlation (ρ=.68, p=5x10-15)

in the two synaptic strengths between cells with bidirectional connections. All networks with a

significant excess of bidirectional connections produced a positive correlation in the strengths of

reciprocal synapses if trained with triplet-STDP. Such a result may be expected from the Hebbian nature

of triplet-STDP and agrees with work using another Hebbian plasticity mechanism (Clopath et al., 2010)

and cortical slice data (Song et al., 2005).

Simplifying the protocol or the network

By simplifying the protocol to be a single stimulus at a time, we increased the specificity and the

sparseness of neural firing. In these examples, the network is not producing any function other than

relaying in a noisy manner the information already present in the input groups, so in our view it is not an

ideal model of cortical activity. Nevertheless the simplified protocol highlights that once we had a

network with different subsets of cells active at different times, our implementation of structural

plasticity enhanced the amount of bidirectional connectivity (Figure 8). Indeed, since selective responses

to single stimuli appeared in the majority of initial, randomly connected networks, structural plasticity

alone was sufficient to produce the observed excess (Figure 8A). Such a result is to be anticipated from

earlier modeling work (Clopath et al., 2010) in which subsets of neurons are specifically excited

together, to produce similar structural changes.

Figure 9A, C indicated that the simplified protocol in our default network (2 input groups per

stimulus with connection probability 1/5) readily reproduced the excess of connected triplets (motifs 10

17
or higher). Moreover, motif 4 was observed at levels significantly higher than chance, in agreement with

slice data (Song et al., 2005).

Finally, we wished to assess how much of the slice data could be reproduced by a simple

network of four cell assemblies. We defined cell assemblies as subsets of neurons with intra-subset

connection probabilities and connection strengths double that of connection probabilities and connection

strengths to other cells (inter-subset) . We found an excess of bidirectional connections (p=3x10-37, ratio

to number expected was 1.4) and that bidirectional connections were on average stronger than

unidirectional connections (p<10-100, ratio was 1.37). Even though connections within a cell assembly

were not correlated, because some bidirectional connections arose between assemblies – and in these

cases both connections would be weak – the network produced a small but highly significant correlation

in the strengths of the two synapses between pairs of cells with a bidirectional connection (ρ = 0.23,

p=2x10-9). The distribution of synaptic strengths was the sum of two overlapping Gaussians (one with

mean 1, the other with mean 2, each with s.d 0.5). Since the Gaussian with higher synaptic strength

contained fewer total connections (only ¼ of all cell pairs were within the same cell assembly, so only ½

of connections were strong and within a cell assembly) the overall shape of synaptic strength

distribution had the same skew as the slice data (Song et al., 2005; Lefort et al., 2009).

Figure 9B,D indicates the patterns of connectivity among triplets of cells. Although absolute

changes in the numbers compared to chance were relatively small (Figure 9B) nearly all patterns were

present at significantly greater than or significantly less than chance (outside the dotted lines denoting

p=0.001 in Figure 9D). In particular, the number of triplets with zero (motif 1), one (motif 2) or three

(motifs 10-16) linked cells was greater than chance, whereas the number of triplets with only two of the

three links present (motifs 4 to 9) was lower than chance. Thus, the only significant aspect of the slice

data missing from the cell assembly structure was the observed excess of motif 4 and of motif 8 (Song et

al., 2005).

18
Discussion

We have shown, in initially randomly connected networks, that the non-random structural

correlations observed in cortical slice data (Song et al., 2005) can arise given appropriate stimuli and

plasticity mechanisms. Intriguingly, when networks are trained to produce a computationally useful

response such as Exclusive-Or (Figure 1A) to paired stimuli, the excesses in numbers of both

bidirectional connections and fully connected triplets appears almost exclusively in those networks that

produce reliable responses to the task, as measured by accumulated reward. Such a correlation between

reward and clustering of connections arises because both require specificity in neural responses – that is,

neurons should respond actively to a small subset of stimuli.

In prior work (Bourjaily, 2010) we have shown the benefit of LTPi in randomly connected

networks for producing specific stimulus-pair neural responses, enabling appropriate behavioral

responses – and thus high accumulation of reward – in a task with Exclusive-Or logic. We chose an

Exclusive-Or logic task since it requires the most basic computation from which any other can be

achieved. Moreover, the accruing of multiple inputs and response to only a specific combination of

active inputs – as measured by paired stimulus selectivity – is a useful feature of information processing,

whereas response to a single input merely relays information, which is hardly a feature of most cortical

circuits. LTPi sculpts the neural response to pairs of inputs by producing cross-inhibition, allowing

distinct subsets of cells to respond to particular input pairs. This increases both sparseness and

specificity of neural responses. Increasing inhibition in a non-specific manner, while increasing the

sparseness of neural responses and reducing overall firing rates, did not allow for strong responses to

specific pairs of inputs. Thus, the presence of an inhibitory plasticity mechanism (LTPi) is essential in

our protocol to produce the correlations in activity among excitatory cells that lead to the

experimentally observed clustering of excitatory-to-excitatory synaptic connections.

19
 Most of the features observed in the slice data – namely high clustering as evinced by an excess

of bidirectional connections and an excess of fully connected triplets (Song et al., 2005) – arise

whenever structural plasticity is Hebbian in form. Indeed, if we assumed correlated cells were more

likely to be connected with each other and their connections were stronger, most in vitro structural

features arose in a model cell-assembly network, in which cells were randomly assigned one of four sub-

groups then assumed to have correlated activity when within the same subgroup. However, when

analyzing the patterns of connections among triplets of cells, two motifs (motif 4 and motif 8) were

significantly depleted, yet observed significantly in excess in the cortical slice data.

The significantly non-random abundances of some motifs place constraints on the

implementations of structural plasticity. In particular, an excess of motif 4 – where one cell projects to

two other non-connected cells – arises when presynaptic partners are replaced upon removal of a

synapse. The mechanism is akin to a shift in dendritic spines until they connect with presynaptic cells

that often produce spikes before the postsynaptic cell. In particular, this process leads to increased

numbers of connections from the most strongly responsive cells to any stimulus towards other cells

active to the same stimulus. Thus, we predict that cells at the “hub” of motifs (such as the upper cell in

motif 4, Figure 3G) are the cells with strongest responses to stimuli that evoke significant local activity.

Ongoing experimental work combining measurements of neural activity with neural connectivity

patterns (Lefort et al., 2009) will provide invaluable data to verify or constrain such predictions that link

plasticity with function.

We implemented a Hebbian-like mechanism for structural plasticity with no homeostatic

component at the single neuron level. Just as Hebbian forms of synaptic plasticity lead to competition in

synaptic strengths – inputs to some cells become stronger than others – so too does the Hebbian form of

structural plasticity lead to competition for numbers of synapses, in particular because in our

implementation the total number of synapses is conserved. Thus, without any structural homeostatic

element, in our networks some cells’ axonal arbors proliferate and some cells’ dendritic arbors
20
proliferate their contacts, while other cells lose contacts (Figure 2). In fact, our inclusion of homeostasis

of synaptic strength (Turrigiano et al., 1998b; Turrigiano, 1999; Turrigiano and Nelson, 2000) but not of

synaptic number at the single neuron level, may have prevented one of the expected results – that

synapses between neurons with bidirectional connections are stronger than those between neurons with

unidirectional connections. We only saw such a result in some of the sparsest networks with few

selectively active neurons that were more often than chance mutually connected and with particularly

strong synapses.

In order to maintain useful function for all cells, we suggest a structural homeostatic element is

necessary at the single neuron level (Kirov and Harris, 1999; Kirov et al., 2004; Butz et al., 2009b; Butz

et al., 2009a), so that the threshold for synaptic loss is increased (making loss harder) among cells with

low activity and decreased (making loss easier) for cells with high activity. Whether such homeostasis

need only act postsynaptically to maintain sufficient inputs and neural activity (Butz and Teuchert-

Noodt, 2006), or whether a presynaptic mechanism is also necessary to ensure sufficient axonal

contacts, is a question for future investigation.

Acknowledgments

We are grateful for financial support from the Swartz Foundation and from NSF via an IGERT award to

the Neuroscience Graduate Program of Brandeis University.

21
Figure Legends

Figure 1: Biconditional Discrimination Task Logic and Network Architecture. A. In an example of

this task, two of four possible stimuli (A, B, C and D) are presented simultaneously to a subject. If either

both A and B are present or neither is present, the subject should make one response (such as release a

lever). If either A or B but not both are present, the subject should make an alternative response (such as

hold the lever until the end of the trial). To perform this task successfully, neurons must generate

responses to specific stimulus-pairs (e.g. A+B). A response to a single stimulus (e.g. A) is not sufficient

to drive the correct response in one pairing without activating the incorrect response for the opposite

pairing of that stimulus. B. The network consists of Poisson input groups that randomly project to a

random recurrent network of excitatory (red) and inhibitory (cells). Excitatory-to-excitatory connections

(arrows) and inhibitory-to-excitatory connections (balls) are probabilistic and plastic. All-to-all

inhibitory-to-inhibitory synapses are also present but not plastic. In the relevant simulations, STDP

occurs at excitatory-to-excitatory and input-to-excitatory synapses, while LTPi occurs at inhibitory-to-

excitatory synapses. Inhibition is feed forward only (i.e. the network does not include recurrent

excitatory-to-inhibitory synapses). C. Excitatory cells from the Associative layer project all-to-all,

initially with equal synaptic strength to excitatory cells in both the hold and release pools of the

decision-making network. The decision-making network consists of two excitatory pools with strong

recurrent connections, which compete via cross-inhibition. Strong self-recurrent excitation ensures

bistability for each pool, while the cross-inhibition generates winner-take-all (WTA) dynamics such that

only one population can be active following the stimulus, resulting in one decision. Whether the motor

output (based on the decision of hold versus release) is correct for the corresponding cue, determines the

presence of Dopamine (DA) at the input synapses, according to the rules of the task in A.

22
Figure 2: Histograms of numbers of outgoing and incoming intra-network connections per cell for

different structural plasticity mechanisms. A) Switching presynaptic cell. B) Switching postsynaptic

cell. C) Select with 50% probability either presynaptic cell or postsynaptic cell to be switched. In all

cases, network with 2 input groups per stimulus, 1/3 input connection probability and undergoing

combined triplet-STDP with LTPi.

Figure 3: Overrepresentation of specific three-cell motifs. A)-C) Ratio of numbers of motifs

produced to numbers expected by chance, given the unidirectional and bidirectional connection

probabilities. D)-F) Z-scores for the numbers of motifs plotted on a non-linear scale. Dashed-lines

represent p = .001, Z = +/- 3.3. A),D) LTPi alone. B),E) Triplet STDP alone. C),F) Combined triplet-

STDP with LTPi. G) List of motifs representing the possible connectivity patterns of three cells (Song et

al., 2005). Motifs numbered 10 or higher are connected triplets.

Figure 4: Inhibitory plasticity boosts numbers of bidirectional connections between excitatory

cells. Sets of five by five networks with different degrees of sparseness of inputs (y-axis where low input

probability reflects high sparseness of inputs) and different degrees of input correlations (x-axis where

more input groups per stimulus reflects lower correlations). (A) Network with no functional plasticity,

just structural plasticity. (B) Network with triplet-STDP and structural plasticity. (C) Network with LTPi

and structural plasticity. (D) Network with triplet-STDP, LTPi and structural plasticity. In both cases

(A) to (C) and (B) to (D) addition of inhibitory plasticity increases the numbers of networks with excess

bidirectional connections. (Color bar, red = twice chance, blue = chance).

Figure 5: Inhibitory plasticity boosts numbers of connected triplets of excitatory cells. Sets of five

by five networks with different degrees of sparseness of inputs (y-axis where low input probability

reflects high sparseness of inputs) and different degrees of input correlations (x-axis where more input
23
groups per stimulus reflects lower correlations). (A) Network with no functional plasticity, just

structural plasticity. (B) Network with triplet-STDP and structural plasticity. (C) Network with LTPi and

structural plasticity. (D) Network with triplet-STDP, LTPi and structural plasticity. In both cases (A) to

(C) and (B) to (D) addition of inhibitory plasticity increases the numbers of networks with excess

connected triplets. (Color bar, red = twice chance, blue = chance).

Figure 6: Excess of bidirectional connections and three-cell connection motifs is highly correlated

with task performance across networks. A)-D) Each data point represents results for a single network.

Only networks with high stimulus selectivity (averaged over cell responses) produce (A) excess

bidirectional connections and (B) excess connected triplets. The cumulative reward, given by the

average reward per trial, is highly correlated with numbers of (C) bidirectional connections and (D)

connected triplets.

Figure 7: Synaptic strengths are not distributed randomly. (A) Histogram of synaptic strengths

reveals a skew, with a long non-Gaussian tail to higher strengths. (B) Synaptic strengths of bidirectional

connections reveal a high correlation between one direction and its reverse. Synapses of all bidirectional

pairs are plotted.

Figure 8: Stronger stimulus specificity produces more networks with high numbers of

bidirectional connections. If single stimuli are activated, rather than paired stimuli, more selective

responses arise, with less need for functional plasticity. Sets of five by five networks with different

degrees of sparseness of inputs (y-axis where low input probability reflects high sparseness of inputs)

and different degrees of input correlations (x-axis where more input groups per stimulus reflects lower

correlations). (A) Network with no functional plasticity, just structural plasticity. (B) Network with

triplet-STDP and structural plasticity. (C) Network with LTPi and structural plasticity. (D) Network
24
with triplet-STDP, LTPi and structural plasticity. While in both cases (A) to (C) and (B) to (D) addition

of inhibitory plasticity increases the numbers of networks with excess bidirectional connections, for

many networks, structural plasticity alone is sufficient to produce excess bidirectional connections.

(Color bar, red = twice chance, blue = chance).

Figure 9: Triplets of connected cells appear with high stimulus specificity and in a cell-assembly

network. A)-B) Ratio of numbers of motifs produced to numbers expected by chance, given the

unidirectional and bidirectional connection probabilities. C)-D) Z-scores for the numbers of motifs

plotted on a non-linear scale. Dashed-lines represent p = .001, Z = +/- 3.3. (A), (C) Network with no

functional, only structural plasticity, but highly selective responses via activation of single stimuli. (B),

(D) Network designed as a set of four cell assemblies, with higher probability of connection within an

assembly versus between assemblies.

25
Figure 1

26
 Figure 2

A B
300
Switch Pre

200

100

C D
300
Switch Post

200

100

E F
Switch Pre/Post

300

200

100

0 50 100 0 50 100
No. of Afferents No. of Efferents

27
 Figure 3

A B C
Count / Expected

100
D E F

10
Z−Score

0
−10
−100

4 10 16 4 10 16 4 10 16

1 2 3 4 5 6 7 8 9

10 11 12 13 14 15 16

28
Figure 4

29
Figure 5

30
 Figure 6

1.5
Mean Selectivity

A B
1

0.5

C D
Mean Reward

0.8
0.7
0.6
0.5
0.4
0.5 1 1.5 2 2.5 3 3.5 0.5 1 1.5 2 2.5 3 3.5
Bidirectional (Relative #) Triplets (Relative #)

31
 Figure 7

A 5000 B

4000
No. of Synapses

0.15

j to i strength
3000

2000 0.1

1000

0.05
0
0 0.1 0.2 0.3 0.4 0.05 0.1 0.15
Synaptic Strength i to j strength

32
Figure 8

33
 Figure 9

A B
Count / Expected

100 C D

10
Z−Score

−10

−100
4 10 16 4 10 16
Triplet Motif No. Triplet Motif No.

34
Alvarez VA, Sabatini BL (2007) Anatomical and physiological plasticity of dendritic spines. Annu
Rev Neurosci 30:79-97.
Baker CI, Behrmann M, Olson CR (2002) Impact of learning on representation of parts and
wholes in monkey inferotemporal cortex. Nat Neurosci 5:1210-1216.
Becker N, Wierenga CJ, Fonseca R, Bonhoeffer T, Nagerl UV (2008) LTD induction causes
morphological changes of presynaptic boutons and reduces their contacts with spines.
Neuron 60:590-597.
Bourjaily M, Miller, P. (2010) Synaptic plasticity and network connectivity requirements to
produce stimulus-pair specific responses in recurrent networks of spiking neurons. PLoS
Comput Biol.
Butz M, Teuchert-Noodt G (2006) A simulation model for compensatory plasticity in the
prefrontal cortex inducing a cortico-cortical dysconnection in early brain development. J
Neural Transm 113:695-710.
Butz M, Worgotter F, van Ooyen A (2009a) Activity-dependent structural plasticity. Brain Res
Rev 60:287-305.
Butz M, van Ooyen A, Worgotter F (2009b) A model for cortical rewiring following
deafferentation and focal stroke. Front Comput Neurosci 3:10.
Cisek P, Puskas GA, El-Murr S (2009) Decisions in changing conditions: the urgency-gating
model. J Neurosci 29:11560-11571.
Clopath C, Busing L, Vasilaki E, Gerstner W (2010) Connectivity reflects coding: a model of
voltage-based STDP with homeostasis. Nat Neurosci 13:344-352.
Dayan P, Abbott LF (2001) Theoretical Neuroscience: MIT Press.
Desimone R, Albright TD, Gross CG, Bruce C (1984) Stimulus-selective properties of inferior
temporal neurons in the macaque. J Neurosci 4:2051-2062.
Dusek JA, Eichenbaum H (1998) The hippocampus and transverse patterning guided by olfactory
cues. Behav Neurosci 112:762-771.
Harris JA, Gharaei S, Moore CA (2009) Representations of single and compound stimuli in
negative and positive patterning. Learn Behav 37:230-245.
Harris JA, Livesey EJ, Gharaei S, Westbrook RF (2008) Negative patterning is easier than a
biconditional discrimination. J Exp Psychol Anim Behav Process 34:494-500.
Holtmaat AJGD, Wilbrecht L, Knott GW, Welker E, Svoboda K (2006) Experience-dependent
and cell-type-specific spine growth in the neocortex. Nature 441:979-983.
Holtmaat AJGD, Trachtenberg JT, Shepherd LWGM, Zhang X, Knott GW, Svoboda K (2005)
Transient and persistent dendritic spines in the neocortex in vivo. Neuron 45:279-291.
Ito M, Tamura H, Fujita I, Tanaka K (1995) Size and position invariance of neuronal responses in
monkey inferotemporal cortex. J Neurophysiol 73:218-226.
Jahr CE, Stevens CF (1990) Voltage dependence of NMDA-activated macroscopic conductances
predicted by single-channel kinetics. \hboxJ Neurosci 10:3178-3182.
Kirov SA, Harris KM (1999) Dendrites are more spiny on mature hippocampal neurons when
synapses are inactivated. Nat Neurosci 2:878-883.
Kirov SA, Goddard CA, Harris KM (2004) Age-dependence in the homeostatic upregulation of
hippocampal dendritic spine number during blocked synaptic transmission.
Neuropharmacology 47:640-648.
Lefort S, Tomm C, Floyd Sarria JC, Petersen CC (2009) The excitatory neuronal network of the
C2 barrel column in mouse primary somatosensory cortex. Neuron 61:301-316.
Maffei A, Nataraj K, Nelson SB, Turrigiano GG (2006) Potentiation of cortical inhibition by visual
deprivation. Nature 443:81-84.

35
Minerbi A, Kahana R, Goldfeld L, Kaufman M, Marom S, Ziv NE (2009) Long-term relationships
between synaptic tenacity, synaptic remodeling, and network activity. PLoS Biol
7:e1000136.
Pfister JP, Gerstner W (2006) Triplets of spikes in a model of spike timing-dependent plasticity. J
Neurosci 26:9673-9682.
Renart A, Song P, Wang XJ (2003) Robust spatial working memory through homeostatic synaptic
scaling in heterogeneous cortical networks. Neuron 38:473-485.
Sakai K, Miyashita Y (1991) Neural organization for the long-term memory of paired associates.
Nature 354:152-155.
Serre T, Kreiman G, Kouh M, Cadieu C, Knoblich U, Poggio T (2007) A quantitative theory of
immediate visual recognition. Prog Brain Res 165:33-56.
Song S, Sjostrom PJ, Reigl M, Nelson S, Chklovskii DB (2005) Highly nonrandom features of
synaptic connectivity in local cortical circuits. PLoS Biol 3:e68.
Toni N, Buchs PA, Nikonenko I, Bron CR, Muller D (1999) LTP promotes formation of multiple
spine synapses between a single axon terminal and a dendrite. Nature 402:421-425.
Trachtenberg JT, Chen BE, Knott GW, Feng G, Sanes JR, Welker E, Svoboda K (2002) Long-
term in vivo imaging of experience-dependent synaptic plasticity in adult cortex. Nature
420:788-794.
Tuckwell HC (1988) Introduction to Theoretical Neurobiology: Cambridge Univ. Press,
Cambridge, U.K.
Turrigiano GG (1999) Homeostatic plasticity in neuronal networks: the more things change the
more they stay the same. Trends Neurosci 22:221-227.
Turrigiano GG, Nelson SB (2000) Hebb and homeostasis in neuronal plasticity. Curr Op Neurobio
10:358-364.
Turrigiano GG, Leslie KR, Desai NS, Rutherford LC, Nelson SB (1998a) Activity-dependent
scaling of quantal amplitude in neocortical neurons. Nature 391:892-896.
Turrigiano GG, Leslie KR, Desai NS, Rutherford LC, Nelson SB (1998b) Activity-dependent
scaling of quantal amplitude in neocortical neurons. Nature 391:892-896.
Wang XJ (2002) Probabilistic decision making by slow reverberation in cortical circuits. Neuron
36:955-968.
Wilbrecht L, Holtmaat A, Wright N, Fox K, Svoboda K (2010) Structural plasticity underlies
experience-dependent functional plasticity of cortical circuits. J Neurosci 30:4927-4932.
Yuste R, Bonhoeffer T (2004) Genesis of dendritic spines: insights from ultrastructural and
imaging studies. Nat Rev Neurosci 5:24-34.

36