Académique Documents
Professionnel Documents
Culture Documents
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
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
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
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
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;
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
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)
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
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
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
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
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
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.
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
Synaptic interactions
Synaptic currents were modeled by instantaneous steps after a spike followed by an exponential
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):
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.
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:
progressively as the number of inputs increased due to the increasing number of independent input
Five levels of input sparseness, combined with five different degrees of input correlations led to
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.
connections are random and uncorrelated. Inhibition is feedforward only, so there are no excitatory-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
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
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
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.
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
(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
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
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
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
> 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.
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
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.
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
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
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
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
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
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
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
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
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,
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
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)
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
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
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,
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
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.
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
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
Acknowledgments
We are grateful for financial support from the Swartz Foundation and from NSF via an IGERT award to
21
Figure Legends
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
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
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
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
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
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
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
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.
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
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