A network of spiking neurons that can represent interval timing: mean field analysis.
ABSTRACT Despite the vital importance of our ability to accurately process and encode temporal information, the underlying neural mechanisms are largely unknown. We have previously described a theoretical framework that explains how temporal representations, similar to those reported in the visual cortex, can form in locally recurrent cortical networks as a function of reward modulated synaptic plasticity. This framework allows networks of both linear and spiking neurons to learn the temporal interval between a stimulus and paired reward signal presented during training. Here we use a mean field approach to analyze the dynamics of nonlinear stochastic spiking neurons in a network trained to encode specific time intervals. This analysis explains how recurrent excitatory feedback allows a network structure to encode temporal representations.

 SourceAvailable from: Daya S Gupta
Dataset: Corticostriatal Clock Circuit
 SourceAvailable from: Daya S Gupta
Page 1
A network of spiking neurons that can represent interval timing:
mean field analysis
Jeffrey P. Gavornik1,2,3 and Harel Z. Shouval1,*
1Department of Neurobiology and Anatomy, The University of Texas at Houston
2Department of Electrical and Computer Engineering, The University of Texas at Austin
Abstract
Despite the vital importance of our ability to accurately process and encode temporal information,
the underlying neural mechanisms are largely unknown. We have previously described a
theoretical framework that explains how temporal representations, similar to those reported in the
visual cortex, can form in locally recurrent cortical networks as a function of reward modulated
synaptic plasticity. This framework allows networks of both linear and spiking neurons to learn
the temporal interval between a stimulus and paired reward signal presented during training. Here
we use a mean field approach to analyze the dynamics of nonlinear stochastic spiking neurons in
a network trained to encode specific time intervals. This analysis explains how recurrent excitatory
feedback allows a network structure to encode temporal representations.
1 Introduction
Disparate visual stimuli can be used as markers for internal time estimates, for example
when determining how long a traffic light will remain yellow. The idea that neurons in the
primary visual cortex might contribute explicitly to this ability contradicts our understanding
of V1 as an immutable visual feature detector and the prevailing notion that temporal
processing is a higherorder cognitive function (Mauk and Buonomano, 2004). These
expectations of V1 function are challenged by the finding that neurons in rat V1 can learn
robust representations of the temporal offset between a visual stimulus and water reward
presented during a behavioral task (Shuler and Bear, 2006). Experimental results suggesting
that that temporal processing might begin in other primary sensory regions have been
reported as well (Super et al., 2001; Moshitch et al., 2006). These observations led us to
investigate how local networks or single neurons can learn, as a function of reward,
temporal representations in lowlevel sensory areas.
In a previous work (Gavornik et al., 2009), outlined below, we demonstrated that recurrent
networks can use reward modulated Hebbian type plasticity as a mechanism to encode time.
Here, we presents a mean field theory (MFT) analysis of temporal representations generated
by a network of conductance based integrate and fire neurons (described in section 3). This
analysis specifically addresses the mechanistic question of how lateral excitation between
nonlinear spiking neurons can be used as the substrate to encode specific durations of time.
We first perform MFT analysis on a noise free system (section 4) then describe and compare
the results of this analysis to those simulated in the full network and show that the temporal
report is invariant to the magnitude of the stimulus (section 5) and that these representations
*To whom correspondence should be addressed, Harel.Shouval@uth.tmc.edu.
3Current Address: The Picower Institute for Learning and Memory, Massachusetts Institute of Technology
NIH Public Access
Author Manuscript
J Comput Neurosci. Author manuscript; available in PMC 2012 April 1.
Published in final edited form as:
J Comput Neurosci. 2011 April ; 30(2): 501–513. doi:10.1007/s108270100275y.
NIHPA Author Manuscript
NIHPA Author Manuscript
NIHPA Author Manuscript
Page 2
can be used to accurately encode short time intervals (section 5.2). Next, we show how the
strength of recurrent connections effects spontaneous activity levels (section 6). Finally, we
describe how the network operates in the superthreshold bistable regions (section 7).
2 A Model of Learned Network Timing
Theoretical studies have shown that a careful tuning of lateral weights can generate neural
networks with attractor states that can possibly account for the neuronal dynamics observed
in association with working memory (Amit, 1989; Seung, 1996; Eliasmith, 2005). Until
recently the potential for local cortical networks to encode temporal instantiations by
learning specific slow dynamics in response to sensory stimuli had not been considered.
Based on the reports of timing activity in rat V1, we proposed a theoretical framework
showing how a network with local lateral excitatory connectivity can learn to represent
temporal intervals as a function of paired stimulus and reward signals (Gavornik et al.,
2009).
Our model consists of a fully recurrent population of neurons receiving stimulating feed
forward projections. This structure (figure 1A) is roughly analogous to the visual cortex
where V1 neurons receive projections from the LGN and interact locally with other V1
neurons. In order to explain how a network can learn such representations we described a
paradigm, called Reward Dependent Expression (RDE), wherein Hebbian plasticity is
modulated by a reward signal paired with feedforward stimulation during training. Briefly,
RDE posits 1. that the action of a reward signal results in long term potentiation through the
permanent expression of activity driven molecular processes, described as “protoweights”
and 2. that ongoing activity in the network inhibits the expressive action of the reward
signal. These assumptions allow RDE to solve a temporal credit assignment problem
associated with the offset between stimulus and reward during early training sessions.
Additionally, the learning rule naturally allows the network to finetune synaptic weights by
preventing additional potentiation as the network nears its target activity level. After training
with RDE, the network responds to specific feedforward stimulation patterns for a period of
time equal to the temporal offset between the stimulus and a paired reward signal presented
during training. Trialtotrial fluctuations in evoked response duration combined with the
highly nonlinear relation between synaptic weights and the network report time (see figure
5A) impose practical limits on the ability of RDE to encode long report periods.
RDE was formulated specifically to explain how plasticity between excitatory neurons in the
visual cortex could encode temporal intervals cued by visual stimulation, but its principles
may apply in other brain regions as well. The temporal representations created by RDE
consist of periods of poststimulus activity whose durations are interpreted to encode neural
instances of “time”. The form of these representations are qualitatively consistent with the
“sustained response” reported in rat V1 (Shuler and Bear, 2006). Our original work
demonstrated that a learned network structure can result in this form of representation using
both a ratebased linear neuron model and nonlinear integrate and fire neurons. Neuronal
activity dynamics in the linear case are purely exponential and easy to analyze (figure 1B).
Activity dynamics in a network of spiking neurons are characterized by a rapid drop
following stimulation to a plateau level, with activity slowly decreasing during the period of
temporal report, and a second drop back to the baseline level at the time of reward (figure
1C). This behavior is quantitatively similar to the experimental data and differs from more
linear looking ramping activity seen in other brain areas. A key component of the previous
work was to demonstrate that RDE allows the network to precisely tune recurrent synapses
to encode specific times; this is import since recurrent network models can be exquisitely
sensitive to synaptic tuning. Notably we have shown that RDE can be used to learn these
times even in a network of stochastic spiking neurons. Although the mechanism responsible
Gavornik and ShouvalPage 2
J Comput Neurosci. Author manuscript; available in PMC 2012 April 1.
NIHPA Author Manuscript
NIHPA Author Manuscript
NIHPA Author Manuscript
Page 3
for encoding time in our model is recurrent excitation, we also demonstrated that RDE
works in the presence of both feedforward and recurrent inhibition. After training, the
average dynamics in networks including inhibition were similar to the dynamics in purely
excitatory networks.
The aim of this paper is to determine quantitively how the spiking network represents time,
and how neuronal nonlinearity shapes the observed form of its dynamics. Understanding
how the excitability of individual spiking neurons can lead to this networklevel dynamical
activity profile, which the linear analysis can not explain, is critical to understanding
mechanistically how temporal representations might form in biological neural networks.
3 Spiking Network Model
The model network consists of a single layer of N = 100 neurons with full excitatory lateral
connectivity (figure 1A). The recurrent layer is assumed to be roughly analogous V1, which
has a large number of synapses with local origin and where extrastriate feedback accounts
for a small percentage of total excitatory current (Johnson and Burkhalter, 1996; Budd,
1998). Recurrent layer neurons are driven by monocular inputs and receive feedforward
projections that are active only during periods of stimulation.
Spiking neurons were simulated with a conductance based integrate and fire model. The
equation governing the subspiking threshold dynamics of the membrane potential, V, of a
single neuron i is:
(1)
where C is the membrane capacitance, and EL and gL are the reversal potential and
conductance associated with the leak current. This equation applies when Vi < Vθ, where Vθ
is the spike threshold. The variable gE,i represents the total excitatory conductance with
current driven by the reversal potential EE. Inhibitory synaptic connections do not contribute
to the formation of temporal representations and are omitted here for the sake of clarity.
Each synapse contributes the product of its activation level and weight to the total
conductance:
(2)
where J is the number of excitatory synapses driving the neuron and sj (t) is the activity level
of synapse j at time t. The synaptic weight variable Wij is used here to indicate that
conductance is determined by all synaptic connections, both feedforward and recurrent. The
subset of W consisting of only the lateral excitatory connections is an N × N matrix L and,
for the sake of this analysis, we will assume homogeneous connectivity.
Synaptic resources are assumed to be finite and saturate following multiple presynaptic
spiking events; maximal transmembrane current occurs when 100% of synaptic resources
are active. Synaptic activation dynamics are modeled independently for each synapse
according to:
Gavornik and ShouvalPage 3
J Comput Neurosci. Author manuscript; available in PMC 2012 April 1.
NIHPA Author Manuscript
NIHPA Author Manuscript
NIHPA Author Manuscript
Page 4
(3)
The synaptic activation level jumps by a fixed percentage, ρ, with each presynaptic spike
and decays with time constant τs. A biological interpretation is that the neurotransmitter
released by each spike binds a fixed percentage of available postsynaptic receptors, and that
bound neurotransmitter dissociates at a constant rate.
Parameters were chosen to be biologically plausible, based on values used in a previous
computational work (Machens et al., 2005). The resting membrane voltage was set to −60
mV, and reversal potentials for excitatory and leak ionic species were −5, and −60 mV
respectively. Spiking occurred when membrane voltage reached a threshold value Vθ = −55
mV. After spikes, the membrane voltage was reset to Vreset = −61 mV and held for a 2 ms
absolute refractory period. The leak conductance was 10e3 μS and membrane capacitance
was set to give a membrane time constant of 20 ms. Each spike is assumed to utilize
approximately 15% of the available synaptic resources (ρ = 1/7) and, as in other models
(Lisman et al., 1998; Compte et al., 2000), synaptic activation decays with a slow time
constant appropriate for NMDA receptor activation dynamics (τs = 80 ms).
After training, a feedforward pulse that drives the network to a high activity state is
sufficient to evoke a report of encoded time (Gavornik et al., 2009). During the stimulation
period, recurrent layer neurons receive random spikes with arrival times drawn from a time
varying Poission distribution chosen to mimic LGN activity (Mastronarde, 1987). In the
original work, each neuron in the recurrent layer also received random spiking input from
independent Poisson processes with intensity levels set to produce a low level of
spontaneous activity in the network (see section 6).
4 Mean field theory analysis
4.1 Extracting the I/O function for a conductance based neuron
The spiking network model is a high dimensional system comprised of orderN coupled
differential equations describing the membrane voltage and synaptic activation dynamics of
all of the neurons and synapses in the network. The MFT approach ignores the detailed
interactions between individual neurons within this large population and instead considers a
single external “field” that approximates the average ensemble behavior. Stochasticity in the
conductance based integrate and fire model described above results from random synaptic
inputs. Accordingly, the approach here will be to replace random synaptic activations and
resultant currents by their mean values and to analyze dynamics in terms of the inputoutput
relationship of a single neuron. This is similar to the approaches that have been used
previously to analyze and solve manybody system problems in various neural networks
(Renart et al., 2003; Amit and Brunel, 1997; Amit et al., 1985). Since the temporal
representation forms in the recurrent layer of our network we will start by analyzing the case
where all of the excitatory input originates from recurrent feedback.
The first requirement for the meanfield analysis is an accurate description of the firing rate,
ν, of the integrate and fire neuron model as a function of synaptic input over the operating
range of a “temporal report”. This relationship can be investigated numerically by driving
the spiking neuron model at a constant rate and simply counting the spikes resulting in a
fixed amount of time. If any of the neuron parameters change, including the strength of
synaptic weights responsible for afferent current, the curve resulting from the numerical
Gavornik and ShouvalPage 4
J Comput Neurosci. Author manuscript; available in PMC 2012 April 1.
NIHPA Author Manuscript
NIHPA Author Manuscript
NIHPA Author Manuscript
Page 5
approach must be regenerated, limiting its usefulness as a tool to understand network
dynamics. An alternative approach is to quantify the I/O curve analytically.
In the spiking neuron model voltage changes at a rate proportional to the total ionic current
(equation 1). The resting membrane potential is set by the reversal potential of the leak
conductance, which is assumed to be constant in time. Excitatory conductance, however, is a
function of random synaptic input. For a single excitatory synapse, the equation governing
synaptic conductances (equation 3) can be rewritten as a stochastic differential equation:
(4)
where X (t) is a random binary spiking process, SE (t) is a random process describing
excitatory synaptic activation, and the other parameters are as defined previously. Assuming
that X (t) is a temporally uncorrelated stationary Poisson process, the expectation of SE (t)
evolves in time as a function of the expectation of X (t) according to the firstorder
differential equation:
(5)
For the Poisson process, E[X (t)] ≡ μ, which is the presynaptic firing frequency driving the
synapse. Defining sE (t) ≡ E[SE] and assuming the initial condition sE (0) = 0, then:
(6)
for t ≥ 0. The resulting steady state value of sE (t), for a constant value of μ, as t → ∞, is:
(7)
Excitatory conductance through a single synapse is the product of the maximal conductance,
defined as the synaptic weight W, and the synaptic resources activation level. That is:
(8)
The expression for average synaptic activation can now be used to write an equation for the
mean conductance of a single synapse independent of time. Assuming that network activity
has been approximately constant long enough to keep the synapse near its steady state value,
which is the case following feedforward stimulation in the network model described in
section 2, equation 8 can be simplified further by replacing sE (t) in the limit with
which results in a constant steadystate excitatory conductance value:
,
(9)
Gavornik and Shouval Page 5
J Comput Neurosci. Author manuscript; available in PMC 2012 April 1.
NIHPA Author Manuscript
NIHPA Author Manuscript
NIHPA Author Manuscript
Page 6
The firing rate of the conductance based neuron model can be estimated analytically as a
function of the mean input current. Equation 1 describes the subthreshold dynamics of the
integrate and fire neuron model, where net current across the membrane is a function of
driving force and conductances associated with the various ionic species. This can be re
written as:
(10)
where the total conductance is gtot = gE (μ) + gL and the reversal potential currents are Irev =
gE (μ)EE + gLEL. The output spike frequency is the inverse of the time, tspike, required for
the voltage to increase from the reset level, Vreset, to the spike threshold, Vthresh, and can be
calculated directly from equation 10 by separating the variables and integrating. The result,
based only on the mean current without fluctuations for a single input spike frequency, is:
(11)
If tspike is real, the corresponding spike frequency is equal to tspike−1; otherwise the spike
frequency is 0. It is now possible to write an analytical function, ϕ (μ, W), relating the output
spike frequency to the mean input spike rate (though the steady state conductance level) and
synaptic weight by combining equations above.
(12)
An upper limit to the output spike frequency is set by the absolute refractory period, tref.
That is, ϕmax = 1/tref.
Figure 2 demonstrates that the spike rate predicted by the MFT analytical ϕ curve agrees
well with numerical estimates of ν over a range of W above some input frequency threshold.
This spiking threshold, which changes as a function of synaptic weight, occurs above the
input where fluctuation driven output is seen in numerical IO solution (see also section 6
and the discussion). Its value is determined by the minimum input current required to drive
the membrane voltage all the way to threshold, which occurs when the numerator in the log
operand of equation 12 is equal to 0. In terms of the excitatory conductance, the threshold
value is:
(13)
and the corresponding input frequency threshold is:
(14)
For a given synaptic weight, spiking will occur whenever μ ≥ μθ. The region above μθ is a
mean driven firing region, in which firing rates are well approximated by the deterministic
Gavornik and ShouvalPage 6
J Comput Neurosci. Author manuscript; available in PMC 2012 April 1.
NIHPA Author Manuscript
NIHPA Author Manuscript
NIHPA Author Manuscript
Page 7
theory (Figure 2A) whereas for input frequency lower than μθ any firing that does occur is
driven by fluctuations from the mean, and cannot be accounted for by the deterministic
approximation.
4.2 Pseudosteady state approximation
Each synapse in the recurrent layer takes a little over 100 ms to reach its steady state
activation level when driven by 50 Hz input. Since the stimulation protocol specifies a
stimulation period of 400 ms, this means that recurrent synapses in the network have
reached approximate steady state activity levels by the beginning of the decay phase.
Furthermore, from equation 6, synaptic activation tracks its steady state value with a time
constant equal to
changes during the temporal report (a decay rate on the order of approximately 1 second).
This implies that a model based on synaptic conductance values equivalent to their steady
state levels as defined in equation 9 should capture decay dynamics during the temporal
report period well.
, which is much faster than the rate that the spike frequency
The formulation of equation 12 assumes only feedforward input and is based on the
relationship between presynaptic spiking activity and excitatory conductance (equations 2
and 4). In the fully recurrent network, however, the recurrent layer’s output is also part of its
own input. Since excitatory current is a function of both the synaptic weight and activation
level, the loop between synaptic activation and spike frequency can be closed by replacing
the generic W component of from equation 12 with the value of the laterally
recurrent weights, L.
The full dynamics of the spiking model are described by equations 1 and 3. Since the
synaptic time constant is assumed to be significantly longer than the effective membrane
time constant, the time course of synaptic activation will dominate membrane voltage
dynamics. This suggests that the differential form of V (equation 1) can be replaced with an
instantaneous function of the average input rate. Accordingly, the dynamics of the recurrent
network model during the falling phase can be described in terms of the synaptic activation
variable, s, and recurrent weight value, L, by replacing the stochastic variable X (t) from
equation 4 with the output frequency calculated using the MFT I/O curve (equation 12):
(15)
where ϕ(s, L) is equivalent to ϕ(μ, W) from equation 12 with
pseudosteady state approximation replaces the full system of equations with a single ODE
and can be used as written or with a numerical estimate of ν replacing the analytical ϕ
function .
replaced by sL. This
5 Dynamics of encoded temporal reports
5.1 Comparison of mean field dynamics to full network dynamics
Equation 15 describes the derivative of synaptic activation as the difference between source
term (ϕ(s, L)ρ(1 − s)), and a negative sink term (s/τs). The relationship between these two
terms as a function of s for different values of the recurrent coupling weight is shown
graphically in figure 3A. In the absence of external input, the system relaxes to a stable
steady state at zero as long as the negative component is larger than the positive component.
The instantaneous decay rate is equal to the the difference between the two components. It is
Gavornik and ShouvalPage 7
J Comput Neurosci. Author manuscript; available in PMC 2012 April 1.
NIHPA Author Manuscript
NIHPA Author Manuscript
NIHPA Author Manuscript
Page 8
immediately clear from this plot that larger recurrent weights increase the response duration
by moving the positive source curve closer to the sink line.
Network dynamics predicted by MFT are found by numerically integrating equation 15 from
an initial condition of s (0) = 1. Dynamics calculated using the deterministic ϕ function for
several values of L are shown in figure 3. These results demonstrate the mechanism
responsible for forming temporal representations; as L increases, the positive and negative
components of equation 15 get closer together, ds/dt gets smaller, and decay dynamics slow
down. A temporal representation results when the recurrent network structure, in effect,
creates a temporal bottleneck in the relaxation dynamics. If the recurrent weights are set too
high, a stable fixed point appears at the upper intersection of the source and sink terms,
corresponding to a high level of persistent firing (see section 7).
The pseudosteady state model explains several features of temporal representations seen in
the spiking network that are absent in the linear model (figure 1). First, instead of decaying
at a constant rate, spiking activity falls quickly to a plateau level following stimulation.
From the MFT analysis it is evident that this results from the large gap, due to synaptic
saturation, between the positive and negative curves at high s values. After the fast initial
drop, activity in the spiking model decays slowly until it reaches some threshold level and
then falls precipitously back to baseline levels. The rapidity of this fall depends on the shape
of the I/O curve (figure 2). A sharp boundary between quiescence and spiking activity, as
predicted by the MFT analysis, will result in a steep drop while the gradual transition seen
using the numerical approach (which includes the noise dominated I/O region) will elicit a
more gradual decay.
The dynamics in figure 3 look qualitatively similar to those seen in the full spiking model.
Figure 4 demonstrates that the dynamics predicted by the pseudo steadystate model for two
values of L accurately describe the dynamics of the full spiking model for the same values of
L (scaled by the number of neurons responsive to the stimulus in the full model) both in
terms of the synaptic activation variable and firing rates. The following relationship, found
by setting equation 15 to zero, is used to convert from synaptic activation levels to steady
state spike frequencies:
(16)
5.2 Limit of encodable time
We can heuristically define the “encoded time” of our network in relation to equation 15.
Here, TE is the time required for the network to relax back to some value close to zero
following stimulation. With this definition relaxation time depends on the value of s at the
end of the stimulus presentation period, but it makes sense to assume a starting point
corresponding to full activation in order to define a measure of the encoded temporal
representation. The existence of an exact solution to equation 15 will depend on the form of
ϕ(s, L), but we can solve for TE by numerically integrating from 1 to some value close to 0.
The result of this calculation is shown in figure 5A.
With the parameters used in this paper, our model is limited to maximum temporal
representations on the order of 1–2 seconds. This is approximately the same response
duration reported by Shuler and Bear, although it is not known wether this duration
represents an upper limit in V1 or is an artifact of the specific stimulusreward offset pairing
presented during training.
Gavornik and Shouval Page 8
J Comput Neurosci. Author manuscript; available in PMC 2012 April 1.
NIHPA Author Manuscript
NIHPA Author Manuscript
NIHPA Author Manuscript
Page 9
5.3 Invariance of temporal report to stimulus intensity
An interesting observation from figure 3 is that the “temporal report” (that is, the duration of
the plateau during which s decays very slowly) occurs over a very narrow range of s values.
Any initial activation greater than the maximal plateau value will report approximately the
same interval. Conversely, any activation below the falloff threshold will report no
temporal representation. This means that the network will reliably report encoded temporal
values so long as stimulation is sufficiently robust to raise the activity level high enough.
This is shown graphically in figure 5B.
There are two implications of this thresholdinvariance that may be of importance in
biological networks: 1. a reliable temporal report requires only a vigorous query of the
trained network and not a carefully graded stimulus 2. subthreshold response dynamics of
individual neurons in the trained network will be no different from those in the naive
network. Since temporal reports require coincident activations of an ensemble of neurons,
temporal representations could conceivably exist on top of other network structures without
changing neural dynamics in the nominal activity range and would not be evident without
the correct querying stimulation pattern.
6 Dynamics and steady state with spontaneous activity
The analysis presented above assumes that activity dynamics during the decay phase are set
only by synaptic connections within the recurrent layer and clearly describes the mechanism
responsible for creating temporal representations in our model and the relationship between
recurrent weights and decay period dynamics. In principle, the same analysis will also
describe changes in steady state firing rates so long as the I/O curve includes an accurate
description of the fluctuation dominated region where spontaneous activity occurs (see
figure 2). In our previous work (Gavornik et al., 2009), spontaneous activity was simulated
by including independent excitatory feedforward synapses into the recurrent population. As
shown in figure 6, training a recurrent network on a timing task increases the rate of
spontaneous activity as the magnitude of the recurrent weights grows even though the
strength of the feedforward synapses driving the activity do not change.
The analytical solution (equation 12) considers activity through a class of synapses with a
single time constant, all of which spike at the same rate. While this approach is is sufficient
to account for the recurrent activity in our network, where all synapses have the same time
constant, it does not capture changes in the low spontaneous firing rate resulting from
synapses with different time constants that spike at a different rate from the recurrent layer
neurons. Although a general analytical solution to this problem is difficult to calculate,
dynamics can be approximated using the pseudosteady state approach by generating
numerical estimates of the I/O curves that include both recurrent and feedforward synapses.
As in the full network model shown in figure 6, spontaneous activity is generated by feed
forward synapses with a time constant τf = 10 ms and maximum conductance of 2.1e2 μS
driven independently by 12 Hz poisson spikes. Dynamics are calculated by replacing the ϕ
function in equation 15 with these numerical estimates of ν as a function of μ.
As seen in figure 7A, the source term is simply proportional to (1 − s) when L = 0. For L >
0, the source curve is much more complicated. This complexity results from the inclusion of
two different time constants associated with the feedforward and feedback components of
excitatory current and will not be further discussed here. Since the feedforward activity
requires a high variability to produce a high CV in the spontaneous activity (cortical neurons
have a CV that is close to 1), the numerical estimates of ν are noisier than those shown in
figure 2. Steady state activity levels are found at the intersection of the source and sink
curves, and in both cases match those seen in the full network (predicted spontaneous rates
Gavornik and Shouval Page 9
J Comput Neurosci. Author manuscript; available in PMC 2012 April 1.
NIHPA Author Manuscript
NIHPA Author Manuscript
NIHPA Author Manuscript
Page 10
of ≈ 4.2 Hz for L = 0 and ≈ 12.5 Hz for L = 3.4e3 μS). Figure 7B demonstrates that the
dynamics calculated using the numerical ν estimate match the dynamics from the full
network model very well.
7 Bistability in the recurrent network
It is clear from figure 3 that the pseudo steady state model has a single fixed point at S = 0
when L is low. If L is increased beyond some threshold value, however, the positive and
negative components of equation 15 will intersect and the system will have three fixed
points. For example, the black source curve in figure 3A intersects the dashed sink line at
three points. In this case the system is bistable and may, depending on the activation level,
move towards a high “up” state (set by the upper intersection) rather than decaying towards
zero (the intersection at the origin). The RDE timing model analyzed in this work requires
the recurrent weights be set below the threshold value so that the system will always decay
to a single fixed point.
The MFT model can be used to demonstrate and analyze the range of bistability as a
function of L. The zeros of equation 15 (that is,
L and holding all other parameters constant. This analysis was repeated over a range of τs
values. The resulting bifurcation diagram is shown in figure 8A. The spike frequencies for
each fixed point value of s were also calculated using equation 16. As expected, the system
is monostable with a solution at s = 0 for low values of L; above some threshold, a second
stable steady state emerges. This “up” state corresponds to persistent firing at a fixed rate,
thought to underlie the persistent activity associated with working memory (Wang, 2001;
Miller et al., 2003). Note that although the lowest possible stable “up” state s value increase
with τs, the minimum possible spike frequency decreases. This accords with previous
findings that slow NMDA currents are critically important in working memory models that
spike at biologically plausible frequencies (Wang, 2001; Seung et al., 2000).
) were found numerically while varying
Bistability can also emerge as a function of other key parameters including τs. As before, a
bifurcation diagram was generated by finding fixed points numerically over a range of τs
with several fixed values of L. The results are shown in figure 8B. Again, the system is
monostable below a threshold and a bistable, with steady solutions at zero and a high “up”
state.
8 Discussion
Our previous work explains qualitatively how temporal representations of the type reported
by (Shuler and Bear, 2006) can form in local recurrent networks as a function of reward
modulated synaptic potentiation (Gavornik et al., 2009). It could not explain the quantitative
form of representations that develops when RDE is applied to a network of spiking neurons.
Specifically, analysis based on a linear neuron model fails to explain why the spike rate falls
so precipitously immediately after feedforward stimulation ends, the rapid drop back to
baseline firing rates at the end of the temporal report, and the increase in spontaneous firing
rates that occurs with training. By reducing the stochastic system of 100 coupled nonlinear
differential equations to a single deterministic ODE, the pseudosteady state model based on
the MFT approach (equation 15) explains these features. Temporal representations form in
networks of nonlinear spiking neurons when the level of recurrent excitatory feedback is
slightly less than the intrinsic neural activity decay rate, resulting in a critical dynamical
slowdown. The particular shape of the temporal representation, as seen in figures 1, 4 and 7
depends on the nonlinear inputoutput relationship of the individual neurons. Unlike in the
Gavornik and ShouvalPage 10
J Comput Neurosci. Author manuscript; available in PMC 2012 April 1.
NIHPA Author Manuscript
NIHPA Author Manuscript
NIHPA Author Manuscript
Page 11
linear case, where activity decays exponentially, spiking temporal representations are
relatively invariant to stimulus magnitude above some threshold (figure 5B).
It should be noted that the analytical MFT approach described here considers only first order
spike statistics. While the meanbased analytical solution of the inputoutput curve matches
the numerically extracted curve well above the spikefrequency threshold (equation 13) it
can not account for the subthreshold region where relatively lowrate spiking is driven by
input fluctuations (figure 2). The agreement between the encoded time predicted by the
analytical meanfield approximation and the full spiking model seen in figure 4 indicates
that the impact of this omission on dynamics is minimal with this parameter set. This can be
understood since the noisedominated region of the ϕ function exists primarily at activation
levels below the narrow bottleneck responsible for the the critical slowing, as seen in figure
3A. Since the analytical solution predicts zero output for low input spike frequencies it can
not explain changes in spontaneous activity levels when each neuron in the recurrent layer is
driven by low levels of random feedforward activity. The pseudosteady state model can
accurately predict these changes if the analytically calculated ϕ function in equation 15 is
replaced by a numerical estimate of the I/O curve that is generated including the feed
forward input (figure 7). The same approach would also work with an analytical I/O curve
including an accurate description of the noise dominated spike region, although the
calculation of such for the conductance based neuron model used here is beyond the scope
of this work.
The CVs of neurons in our network are close to 1 when spontaneously active and drop to a
value closer to 0.4 following stimulation. A similar phenomenon, whereby external stimulus
onset decreases neural spiking variability, has been reported in various brain regions
(Churchland et al., 2009). The CVs of V1 neurons during the temporal report period have
not been experimentally characterized, although they are likely to be higher than exhibited
in our spiking network. Computational work has shown that balanced network models
including high reset values and short term synaptic depression or neuronal adaptation can
exhibit high CVs concurrent with stored memory retrieval and bistable “up” states (Barbieri
and Brunel, 2008; Roudi and Latham, 2007). It is an open question how the inclusion of the
additional mechanisms used in these models to increase variance would impact our timing
model, which depends on the development of regular slow dynamics near bifurcation points.
Figure 8 demonstrates that the mechanism underlying RDE has the potential to create
regions of bistability. In the bistable regime, our model becomes very similar to models of
persistent activity thought to underlie working memory processes(Lisman et al., 1998;Wang,
2001;Seung, 1996;Brody et al., 2003;Miller et al., 2003). A learning rule similar to RDE
could be used to tune a recurrent network to produce desired “up” levels.
It has been suggested that temporal processing might involve the same mechanisms that
underlie working memory (Lewis and Miall, 2006; Staddon, 2005), and correlates of
working memory have been observed in the monkey visual cortex (Super et al., 2001).
Despite this, there is no experimental evidence that the high stable state is reached, or used,
by the neurons in V1. Functionally, it makes little sense that persistent firing in the low level
visual processing areas to result from a brief stimulus would be desirable to the visual
system. Presumably, homeostatic mechanisms not included in our model could ensure that
the high state is never reached in V1. The bifurcation analysis helps set upper limits on the
allowable range of parameters in the recurrent network over which the model of temporal
representation is valid. Based on the similarity between our RDE model of temporal
representation and previous models of working memory, it is possible that similar neural
machinery is recruited for both tasks by the brain.
Gavornik and ShouvalPage 11
J Comput Neurosci. Author manuscript; available in PMC 2012 April 1.
NIHPA Author Manuscript
NIHPA Author Manuscript
NIHPA Author Manuscript
Page 12
Our model, as presented and analyzed in this work, contains no role for inhibition. Indeed,
RDE posits that recurrent excitation provides the neural basis for temporal representations.
We have demonstrated that recurrent excitation can overcome both feedforward and
recurrent inhibition and that RDE is able to entrain temporal representations in networks that
include biologically realistic ratios of excitation and inhibition (Gavornik, 2009). Recurrent
network models including significant amounts of inhibition have been successfully analyzed
using the mean field approach (Brunel and Wang, 2001; Renart et al., 2003, 2007).
Another possible way to explain the data in (Shuler and Bear, 2006) is that temporal
representations form within individual neurons independent of network structure. In an
accompanying work (Shouval and Gavornik, 2010), we present a model demonstrating how
single neurons can learn temporal representations through reward based modulation of their
intrinsic membrane conductances. The mechanistic basis of the prolonged spiking in this
alternate model is a positive feedback loop between voltage gated calcium channels and
calcium dependent cationic channels. Although analysis of the single cell model reveals
mathematical similarities with the mean field approach presented here, there are functional
differences between the two models that can be explored experimentally. It is also possible
that temporal representations in the brain could form through a hybridization of the two
models.
The success of RDE in explaining how temporal representations can arise in local cortical
networks, such as V1, suggests that neural mechanisms responsible for temporal processing
may be more distributed throughout the brain than previously thought. Additional reports of
persistent activity in other sensory cortices may further this hypothesis. The mechanistic
similarity between the temporal representations predicted by RDE and those thought to
contribute to working memory could imply that a simple mechanism, widely available
throughout the cortex, can be recruited for widely different tasks. This work demonstrates
that a MFT approach can explain how temporal representations can form in networks of
nonlinear spiking neurons.
References
Amit, DJ. Modeling brain function : the world of attractor neural networks. Cambridge [England]:
Cambridge University Press; 1989. New York, 89015741 Daniel J. Amit. ill. ; 24 cm. Includes
bibliographies and index
Amit DJ, Brunel N. Model of global spontaneous activity and local structured activity during delay
periods in the cerebral cortex. Cereb Cortex. 1997; 7(3):237–252. [PubMed: 9143444]
Amit DJ, Gutfreund H, Sompolinsky H. Spinglass models of neural networks. Phys Rev A. 1985;
32(2):1007–1018. [PubMed: 9896156]
Barbieri F, Brunel N. Can attractor network models account for the statistics of firing during persistent
activity in prefrontal cortex? Front Neurosci. 2008; 2(1):114–122. [PubMed: 18982114]
Brody CD, Romo R, Kepecs A. Basic mechanisms for graded persistent activity: discrete attractors,
continuous attractors, and dynamic representations. Curr Opin Neurobiol. 2003; 13(2):204–211.
[PubMed: 12744975]
Brunel N, Wang XJ. Effects of neuromodulation in a cortical network model of object working
memory dominated by recurrent inhibition. J Comput Neurosci. 2001; 11(1):63–85. [PubMed:
11524578]
Budd JM. Extrastriate feedback to primary visual cortex in primates: a quantitative analysis of
connectivity. Proc Biol Sci. 1998; 265(1400):1037–1044. [PubMed: 9675911]
Churchland MM, Yu BM, Cunningham JP, Sugrue LP, Cohen MR, Corrado GS, Newsome WT, Clark
AM, Hosseini P, Scott BB, Bradley DC, Smith MA, Kohn A, Movshon JA, Armstrong KM, Moore
T, Chang SW, Snyder LH, Lisberger SG, Priebe NJ, Finn IM, Ferster D, Ryu SI, Santhanam G,
Sahani M, Shenoy KV. Stimulus onset quenches neural variability: a widespread cortical
phenomenon. Nat Neurosci. 2009; 13(3):369–378. [PubMed: 20173745]
Gavornik and ShouvalPage 12
J Comput Neurosci. Author manuscript; available in PMC 2012 April 1.
NIHPA Author Manuscript
NIHPA Author Manuscript
NIHPA Author Manuscript
Page 13
Compte A, Brunel N, GoldmanRakic PS, Wang XJ. Synaptic mechanisms and network dynamics
underlying spatial working memory in a cortical network model. Cereb Cortex. 2000; 10(9):910–
923. [PubMed: 10982751]
Eliasmith C. A unified approach to building and controlling spiking attractor networks. Neural
Comput. 2005; 17(6):1276–1314. [PubMed: 15901399]
Gavornik, JP. Ph.D. dissertation. The University of Texas at Austin; 2009. Learning temporal
representations in cortical networks through reward dependent expression of synaptic plasticity.
Gavornik JP, Shuler MG, Loewenstein Y, Bear MF, Shouval HZ. Learning reward timing in cortex
through reward dependent expression of synaptic plasticity. Proc Natl Acad Sci U S A. 2009;
106(16):6826–6831. [PubMed: 19346478]
Johnson RR, Burkhalter A. Microcircuitry of forward and feedback connections within rat visual
cortex. J Comp Neurol. 1996; 368(3):383–398. [PubMed: 8725346]
Lewis PA, Miall RC. A right hemispheric prefrontal system for cognitive time measurement. Behav
Processes. 2006; 71(2–3):226–234. [PubMed: 16434151]
Lisman JE, Fellous JM, Wang XJ. A role for NMDAreceptor channels in working memory. Nat
Neurosci. 1998; 1(4):273–275. [PubMed: 10195158]
Machens CK, Romo R, Brody CD. Flexible control of mutual inhibition: a neural model of two
interval discrimination. Science. 2005; 307(5712):1121–1124. [PubMed: 15718474]
Mastronarde DN. Two classes of singleinput Xcells in cat lateral geniculate nucleus. II. Retinal
inputs and the generation of receptivefield properties. J Neurophysiol. 1987; 57(2):381–413.
[PubMed: 3559685]
Mauk MD, Buonomano DV. The neural basis of temporal processing. Annu Rev Neurosci. 2004;
27:307–340. [PubMed: 15217335]
Miller P, Brody CD, Romo R, Wang XJ. A recurrent network model of somatosensory parametric
working memory in the prefrontal cortex. Cereb Cortex. 2003; 13(11):1208–1218. [PubMed:
14576212]
Moshitch D, Las L, Ulanovsky N, BarYosef O, Nelken I. Responses of neurons in primary auditory
cortex (A1) to pure tones in the halothaneanesthetized cat. J Neurophysiol. 2006; 95(6):3756–
3769. [PubMed: 16554513]
Renart, A.; Brunel, N.; Wang, X. Meanfield theory of irregularly spiking neuronal populations and
working memory in recurrent cortical networks. In: Feng, J., editor. Computational Neuroscience:
A Comprehensive Approach. Boca Raton: CRC Press; 2003. p. 431490.
Renart A, MorenoBote R, Wang XJ, Parga N. Meandriven and fluctuationdriven persistent activity
in recurrent networks. Neural Comput. 2007; 19(1):1–46. [PubMed: 17134316]
Roudi Y, Latham PE. A balanced memory network. PLoS Comput Biol. 2007; 3(9):1679–1700.
[PubMed: 17845070]
Seung HS. How the brain keeps the eyes still. Proc Natl Acad Sci U S A. 1996; 93(23):13339–13344.
[PubMed: 8917592]
Seung HS, Lee DD, Reis BY, Tank DW. Stability of the memory of eye position in a recurrent
network of conductancebased model neurons. Neuron. 2000; 26(1):259–271. [PubMed:
10798409]
Shouval HZ, Gavornik JP. A single cell with active conductances can learn timing and multistability.
Journal of Computational Neuroscience. 2010 in press.
Shuler MG, Bear MF. Reward timing in the primary visual cortex. Science. 2006; 311(5767):1606–
1609. [PubMed: 16543459]
Staddon JE. Interval timing: memory, not a clock. Trends Cogn Sci. 2005; 9(7):312–314. [PubMed:
15953755]
Super H, Spekreijse H, Lamme VA. A neural correlate of working memory in the monkey primary
visual cortex. Science. 2001; 293(5527):120–124. [PubMed: 11441187]
Wang XJ. Synaptic reverberation underlying mnemonic persistent activity. Trends Neurosci. 2001;
24(8):455–463. [PubMed: 11476885]
Gavornik and Shouval Page 13
J Comput Neurosci. Author manuscript; available in PMC 2012 April 1.
NIHPA Author Manuscript
NIHPA Author Manuscript
NIHPA Author Manuscript
Page 14
Figure 1.
Temporal representations created by RDE. A. Neurons in the recurrent layer of our network
model are stimulated by retinal activation via the LGN. L is the matrix defining lateral
excitation. B. With a linear neuron model, time is encoded by the exponential decay rate of
an activity variable V. C. In the spiking neuron model, evoked activity (shown by spike
rasters, where each row represents a single neuron in the network, and the resultant
histogram) in a responsive subpopulation of the network persists until the time of reward. In
both models, the stimulus is active during the period marked by the gray bar and the reward
time is indicated by the dashed line. See (Gavornik et al., 2009) for details of learning with
RDE.
Gavornik and ShouvalPage 14
J Comput Neurosci. Author manuscript; available in PMC 2012 April 1.
NIHPA Author Manuscript
NIHPA Author Manuscript
NIHPA Author Manuscript
Page 15
Figure 2.
InputOutput relationship of an IF neuron. A. The analytical ϕ curve (black line) calculated
using MFT analysis (equation 12) with W = 3.4e3 μS compared to numerically generated
estimates of the output rate ν (symbols). K indicates the number of independent synapses
driving activity in the numerically simulated neuron; individual synaptic weights are scaled
by K so that the cumulative synaptic weight is constant for each of the three cases shown
(K=1,10,100). As K increases, the numerical approximations approach the analytical curve.
Note that in the model described in section 3, K=N=100 for the recurrent synapses.
Deviations exist primarily in the low frequency input region where output is driven by
fluctuations (see equations 13 and 14). B. The analytical solution (solid lines) compares well
with numerical results (plus signs, K=100) for values of W ranging from 1.5e3 μS (light
gray) to 6.0e3 (black). All parameters are as listed in section 3.
Gavornik and ShouvalPage 15
J Comput Neurosci. Author manuscript; available in PMC 2012 April 1.
NIHPA Author Manuscript
NIHPA Author Manuscript
NIHPA Author Manuscript
Similar Publications
Jeffrey P Gavornik 