(0.2 Mb PDF)

(0.2 Mb PDF)
Neuron, Vol. 35, 773–782, August 15, 2002, Copyright 2002 by Cell Press
Gain Modulation from Background Synaptic Input
Frances S. Chance,1,3 L.F. Abbott,2
and Alex D. Reyes1
1
Center for Neural Science
New York University
New York, New York 10003
2
Volen Center for Complex Systems and
Department of Biology
Brandeis University
Waltham, Massachusetts 02454
Summary
Gain modulation is a prominent feature of neuronal
activity recorded in behaving animals, but the mechanism by which it occurs is unknown. By introducing a
barrage of excitatory and inhibitory synaptic conductances that mimics conditions encountered in vivo into
pyramidal neurons in slices of rat somatosensory cortex, we show that the gain of a neuronal response to
excitatory drive can be modulated by varying the level
of “background” synaptic input. Simultaneously increasing both excitatory and inhibitory background
firing rates in a balanced manner results in a divisive
gain modulation of the neuronal response without appreciable signal-independent increases in firing rate
or spike-train variability. These results suggest that,
within active cortical circuits, the overall level of synaptic input to a neuron acts as a gain control signal
that modulates responsiveness to excitatory drive.
Introduction
Neurons in vivo are continuously bombarded by synaptic input, which dramatically affects their response properties (Destexhe and Paré, 1999; Hô and Destexhe, 2000;
Tiesinga et al., 2001; Anderson et al., 2000a) by increasing overall conductance (Borg-Graham et al., 1998;
Hirsch et al., 1998; Destexhe and Paré, 1999) and introducing a high degree of response variability (Softky and
Koch, 1993; Holt et al., 1996; Stevens and Zador, 1998;
Shadlen and Newsome, 1994; Troyer and Miller, 1997).
Typically, this background activity is treated as a constant source of noise that continuously underlies the
stimulus-evoked increases in excitation that drive neuronal responses. Here we consider the impact of varying
the level of background activity. We find that changing
the level of background input, rather than affecting response variability, modulates the gain of neuronal responses. Our results suggest that the gain modulation
commonly seen in vivo may arise from varying levels of
background synaptic input.
Gain modulation is a primary mechanism by which
cortical neurons combine and process information (for
a review, see Salinas and Thier, 2000). It appears in
a wide range of contexts, including the gaze-direction
dependence of visual neurons in posterior parietal cor3
Correspondence: [email protected]
tex (Andersen and Mountcastle, 1983; Andersen et al.,
1985), the effects of attention (McAdams and Maunsell,
1999a; Treue and Martı́nez-Trujillo, 1999), and as a possible basis for a variety of “nonclassical” receptive field
effects in primary visual cortex (Heeger, 1992). Gain
modulation has also been proposed as a mechanism for
the neural computation of coordinate transformations
relevant for tasks ranging from visually guided reaching
(Zipser and Andersen, 1988; Salinas and Abbott, 1995;
Pouget and Sejnowski, 1997) to invariant object recognition (Salinas and Abbott, 1997).
Gain modulation is not equivalent to the enhancement
or suppression of neuronal responses by pure excitation
or inhibition. To illustrate the distinction, consider the
firing rate of a neuron in response to injected current
(which we call driving current) shown schematically in
Figure 1A. Increased excitation shifts the firing-rate
curve to the left, and increased inhibition, whether of
the hyperpolarizing or shunting variety (Holt and Koch,
1997), shifts it to the right (as discussed below, see
Figure 3A). Gain modulation, on the other hand, is a
change in the slope of the firing-rate curve, corresponding to a multiplicative or divisive scaling, which is distinct
from these additive or subtractive shifts. Mechanisms
that generate true gain modulation from fast, ionotropic
synaptic input have proven elusive (Srinivasan and Bernard, 1976; Koch and Poggio, 1992; Mel, 1993; Salinas
and Abbott, 1996; Hahnloser et al., 2000; Doiron et al.,
2001; Smith et al., 2002).
The background synaptic input we study is primarily in
the balanced configuration proposed to exist in cortical
circuits (Shadlen and Newsome, 1994; Troyer and Miller,
1997). In this configuration, excitatory and inhibitory
components approximately cancel, keeping the average
total synaptic current near zero. Although a balanced
configuration is not a strict requirement for the results
we present, it has the advantage that the level of background activity can be modified independently of excitatory synaptic drive if excitatory and inhibitory inputs
are modulated in parallel, specifically if their input rates
are scaled by the same factor. Conversely, the overall
level of excitatory drive can be modified without changing the total synaptic input to a neuron if excitatory and
inhibitory rates are modulated in an opposing or pushpull manner (Anderson et al., 2000b). Thus, the excitatory
drive and the total level of synaptic input can be modulated independently and comprise two separate input
channels. Traditionally, the push-pull channel has been
considered to be the primary information conduit to the
neuron. Here, we find that the “noise” channel, consisting of the overall level of synaptic input, can carry a
second, independent control signal that modulates the
gain of neuronal responses.
Results
Introducing In Vivo Synaptic Input
into In Vitro Neurons
To show how background synaptic input modulates the
gain of neuronal responses, we introduced conductance
Neuron
774
Figure 1. Simulating Background Synaptic
Input with the Dynamic Clamp
(A) The thick solid trace represents a hypothetical plot of the firing rate of a neuron as
a function of injected current. Adding excitation shifts the curve to the left, and inhibition
(whether hyperpolarizing or shunting) shifts it
to the right (dashed traces). Gain modulation,
on the other hand, corresponds to a tipping
of the firing-rate curve resulting in a change
of its slope (thin solid trace).
(B) The dynamic clamp was used to simulate
in vivo background synaptic conductances
in vitro. Simulated excitatory and inhibitory
synaptic conductances (gE and gI) were generated from independent Poisson spike
trains. The synaptic current, Isyn, was computed by multiplying these computer-generated conductances by the difference between the measured membrane potential and
the synaptic reversal potentials (equation in
box). The computed synaptic current was
then injected into the neuron along with an
additional constant driving current.
(C) Membrane potential recorded from a neuron firing in response to a constant driving
current without (left) and with (right) mimicked
background synaptic input.
fluctuations, mimicking in vivo conditions, into neurons
recorded in vitro. Neurons in a slice preparation receive
little synaptic input because of the general lack of spontaneous activity in the slice. To study the effects of
background input in a slice preparation, we used a computer-controlled dynamic clamp (see Experimental Procedures) to mimic, in single in vitro neurons, the conductance changes caused in vivo by the firing of populations
of excitatory and inhibitory presynaptic neurons (Figure
1B) (see also Destexhe et al., 2001). In brief, the computer simulated the firing of excitatory and inhibitory
afferent populations using independent random Poisson
processes. With each excitatory or inhibitory presynaptic action potential, a unitary conductance, equivalent
to that mediated by either AMPA or GABAA receptors
(see, for example, Destexhe and Paré, 1999), was added
to running tallies of total excitatory or inhibitory synaptic
conductance, respectively. These conductances were
then introduced into the cell under dynamic clamp via
two electrodes patched on the soma.
To simulate typical in vivo conditions, excitatory inputs were generated at a rate of 7000 Hz and inhibitory
inputs at a rate of 3000 Hz, representing the summed
effects of many simulated afferents. We refer to this as
the 1X condition. For the synaptic conductances and
reversal potentials we used, these rates imply that the
average total synaptic current reverses at ⫺57 mV, and
that excitatory and inhibitory contributions to the total
synaptic current approximately cancel each other at typical membrane potentials. Because the excitatory and
inhibitory synaptic inputs that make up the background
synaptic activity were in such a balanced configuration,
high levels of input noise and shunting could be introduced without producing excessive hyperpolarization
or depolarization. We use the term input noise to mean
specifically the variance of the total synaptic input current, and shunting to mean the change in total conduc-
tance due to the combined effects of excitatory and
inhibitory synaptic inputs. The input rates and sizes of
the unitary synaptic conductances were chosen so that
the resulting barrage of background synaptic input affected the neuron in a manner consistent with in vivo
measurements: shunting by background synaptic input
increased the conductance of the neuron by two to three
times its resting value (Bernander et al., 1991; Rapp et
al., 1992; Borg-Graham et al., 1998; Hirsch et al., 1998;
Destexhe and Paré, 1999), and the associated fluctuations in the total synaptic current (the input noise) induced fluctuations with an amplitude of a few millivolts
in the membrane potential (Anderson et al., 2000a),
changing the pattern of spiking in response to injected
driving current from regular to irregular (Figure 1C).
The irregularity of spiking and general response variability introduced by the dynamic clamp input (see Figures 1C, 3E, and 3F) is toward the low end of the range
measured in vivo. This level of variability is roughly the
limit of what can be achieved with reasonable unitary
synaptic conductances driven by uncorrelated inputs,
which is what we used. More variable responses can
be obtained by using correlated synaptic input (Stevens
and Zador, 1998; Hô and Destexhe, 2000), but we decided against this both because of uncertainties about
the amount and nature of the correlations that exist in
vivo, and because the gain modulation effect we report
is most robust when variability is high (discussed later;
and see Figure 4). Thus, the conservative levels of variability produced by uncorrelated synaptic input provide
a more rigorous test of the mechanism.
Multiplicative Gain Modulation of Neurons In Vitro
The point of our study is not only to introduce realistic
background synaptic input into the quiet environment
of the slice but also to see how varying its level affects
neuronal responses. We modified the level of back-
Gain Modulation from Background Synaptic Input
775
ground synaptic input by scaling the computer-generated excitatory and inhibitory input rates by the same
factor, which we call the rate factor (going to 2X or 3X
conditions refers to doubling or tripling both input rates,
for example). This manipulation increases both the input
noise (the variance of the synaptic current) and the
amount of shunting (the average total synaptic conductance) in proportion to the rate factor. Specifically, the
average amount of shunting for excitatory and inhibitory
inputs with rates rE and rI, peak unitary conductances
gE and gI, synaptic decay constants ␶E and ␶I, and reversal
potentials EE and EI is given by gE␶ErE ⫹ gI␶IrI, and the
variance of the fluctuations in the total synaptic current
at membrane potential V is proportional to gE2␶ErE(V ⫺
EE)2 ⫹ gI2␶IrI(V ⫺ EI)2. Because the background input is
in a balanced configuration, equal scaling of excitatory
and inhibitory input rates produces little net hyperpolarization or depolarization.
To characterize neuronal responses, we drove the
neuron by injecting constant current, which we call driving current, along with the mimicked background synaptic input. We also performed studies in which neurons
were driven by dynamic-clamp-generated excitatory
synaptic inputs, rather than by injected current. Additional excitatory synaptic input introduces noise and
shunting that are negligible in comparison to that produced by the balanced synaptic input. As a result, excitatory synaptic drive produces effects that are indistinguishable from those of constant injected current. For
the experiments reported here, we used injected current
to drive the neurons because this allows us to separate
clearly the effects of excitatory drive from those of background synaptic input.
Figure 2A shows firing rates evoked by different
amounts of driving current for various levels of background synaptic input (0X, closed diamonds; 1X, open
circles; 2X, closed squares; 3X, open triangles). Increasing the level of background synaptic input made the
neurons progressively less sensitive to changes in driving current, equivalent to modulating the gain of the
neuronal response. The gains corresponding to the results in Figure 2A are plotted as a function of firing rate
in Figure 2B (0X, closed diamond; 1X, open circle; 2X,
closed square; 3X, open triangle). These gains were determined by taking the derivative of the best secondorder polynomial fit to the firing-rate data. We plot the
gain as a function of firing rate rather than driving current
because this allows us to distinguish an additive (leftright) shift in a nonlinear firing-rate curve from a gain
modulation. An additive shift in the firing-rate curve produces no change in a plot of gain versus firing rate,
whereas a multiplicative modulation produces changes
like those seen in Figure 2B.
We also determined average gains (rather than ratedependent gains) by fitting straight lines to the nonzero
portions of the firing-rate curves obtained using different
levels of background input. The average gains, defined
as the slopes of these straight-line fits, provide a single
measure of response sensitivity. Figure 2C illustrates
such linear fits and also shows gain modulation by background synaptic input in a different neuron than that
shown in Figure 2A. The slopes of the fitted lines in
Figure 2C are 34.4, 25.2, 17.5, and 12.7 Hz/nA in the 0X
(closed diamonds), 1X (open circles), 2X (closed
Figure 2. Changing the Level of Background Input Modulates Gain
(A) Firing rates of a representative neuron as a function of constant
driving current without simulated background synaptic input (0X,
closed diamonds), and with 1X (open circles), 2X (closed squares),
and 3X (open triangles) background synaptic input. The solid lines
are the best second-order polynomial fits to the data.
(B) The gains for the firing-rate curves in (A), defined as the derivatives of the polynomial fitting curves, plotted as a function of firing
rate. Each trace is labeled by the symbol representing the corresponding case in (A).
(C) Firing rates of another neuron as a function of constant driving
current under the same conditions as in (A) (0X, closed diamonds;
1X, open circles; 2X, closed squares; and 3X, open triangles). The
solid lines are the best linear fits to the nonzero portions of the
firing-rate curves.
(D) Average normalized gains for 18 different neurons as a function
of the rate factor. Here, normalized gain is defined as the slope of
the best linear fit to the nonzero portion of the firing-rate curve
divided by the slope for a rate factor of one. Error bars are standard
errors, and the numbers over each data point indicate the number
of neurons tested in each condition.
(E) The firing-rate curves in (A) modified by appropriate scale factors
to align them, with symbols as in (A). This indicates that gain modulation by background synaptic input is multiplicative. The 1X case
appears unchanged from (A), but the driving currents for the 0X,
2X, and 3X data points were scaled by the factors 2, 0.65, and 0.45,
respectively, while the firing rates were left unchanged.
(F) Example of a neuron for which the slope change in the firingrate curve, due to changing the level of background synaptic input
(0X, closed diamonds; 1X, open circles; and 2X, closed squares), is
accompanied by a shift in the firing-rate curves.
squares), and 3X (open triangles) cases, respectively.
Similar fits to Figure 2A yield 28.8, 22.5, 14.3, and 10.8
Hz/nA.
Neuron
776
We studied gain modulation effects due to background synaptic input in 18 neurons, 15 of which showed
consistent decreases in gain as the background input
was increased across all levels tested. In Figure 2D, we
characterize the gain of each neuron in each condition
by the slope of the best straight-line fit to the corresponding firing-rate curve (as in Figure 2C) normalized
by the gain of the neuron in the 1X condition. On average,
changing the level of background input by rate factors
ranging from 0 to 3 produced gain modulation by more
than a factor of two, and this effect extended up to firing
rates of at least 100 Hz. In these experiments, the same
parameters were used for each neuron (see Experimental Procedures), which indicates that the mechanism of
gain modulation does not require fine tuning.
Gain modulation by background synaptic input can
be quite accurately multiplicative, as illustrated in Figure
2E, where the data points from Figure 2A have been
scaled by appropriate factors to make them line up. This
agrees with observations in a variety of in vivo systems
where gain modulation is also multiplicative (Andersen
and Mountcastle, 1983; Andersen et al., 1985; McAdams
and Maunsell, 1999a; Treue and Martı́nez-Trujillo, 1999;
Peña and Konishi, 2001). The firing-rate curves in Figure
2E were aligned by applying scaling factors to the driving
currents, but similar results can be obtained by scaling
the firing rates instead. This is because the firing-rate
curves are approximately described by a power-law
over the range we examine (Miller and Troyer, 2002;
Hansel and Van Vreeswijk, 2002). Gain modulation can
have different functional consequences depending on
the nature of the scaling involved (Reynolds et al., 2000).
We return to this issue in a later section.
For the fixed set of parameters we used in generating
the background synaptic input, some of the neurons
showed gain modulation that was not purely multiplicative because changing the rate factor shifted the firingrate curves as well as changing their slopes. This shifting
effect occurred at least in part because the parameters
we used introduced a slight excess of inhibition. An
example of such a shift, which is at the large end of the
range for neurons showing significant gain modulation
by background synaptic input, is shown in Figure 2F (0X,
closed diamonds; 1X, open circles; 2X, closed squares).
Purely multiplicative gain modulation could be obtained
for this neuron by adjusting the balance between the
inhibitory and excitatory components that make up the
background synaptic input (see below), but we did not
do this because we wanted to avoid fine tuning of parameters for each of the neurons studied.
Contributions of Shunting and Noise
to Gain Modulation
Increasing the rate of background synaptic input increases both the level of input noise and the amount of
shunting. Because shunting produces a change in the
input resistance (R ), which affects the relationship between input current (I ) and membrane potential (V ) in a
multiplicative manner (V ⫽ IR), it is often assumed that
multiplicative gain modulation of neuronal responses
can be generated purely by shunting. However, a theoretical study by Holt and Koch (1997) shows that this
assumption is incorrect in a variety of neuron models.
Shunting generates a shift rather than a slope change
in the firing-rate curve, similar to what occurs with hyperpolarizing inhibition (Figure 1A). For example, the firing
rate of an integrate-and-fire model, for sufficiently high
rates, is approximately proportional to (A ⫹ IR)/␶m, where
A is a constant and ␶m is the effective membrane time
constant (see, for example, Dayan and Abbott, 2001).
This shows the expected dependence on the depolarization IR, but also includes a dependence on the membrane time constant ␶m, because this parameter sets the
time scale for the dynamics of the model. Substituting
the definition ␶m ⫽ RC into the above expression yields
a firing rate proportional to (A ⫹ IR)/RC ⫽ A/(RC) ⫹ I/C.
The first term in this expression produces an additive
shift in the firing-rate curve when R is modified by
shunting, but the second term, which determines the
dependence of the firing rate on the input current, is
independent of R and is thus unaffected by shunting.
This explains why increased shunting shifts rather than
tips firing-rate curves; decreasing R leads to less depolarization (IR), but this effect is cancelled by a decrease
in the membrane time constant (RC).
To determine whether the theoretical arguments given
above and in Holt and Koch (1997) apply to the neurons
we studied, we used the dynamic clamp to introduce a
shunting conductance equivalent to that introduced by
the 2X background synaptic input (32 nS) used in Figures
1 and 2, but without any accompanying input noise. The
reversal potential for this constant conductance was
set equal to the conductance-weighted average of the
excitatory and inhibitory reversal potentials of the background synaptic input (⫺57 mV). In agreement with the
theoretical results, we found that increased shunting
shifted the firing-rate curve but did not change its slope
(n ⫽ 8; for an example see Figure 3A, control [closed
circles], with additional conductance [open squares]).
Increasing the rate of background synaptic input increases the input noise as well as the amount of
shunting. Because shunting alone does not change the
slope of the firing-rate curve, we next consider the role
that input noise plays in the gain modulation effect. To
examine the effects of noise alone, we used the dynamic
clamp to change the level of input noise without modifying the amount of synaptic shunting. This is done by
multiplying the unitary synaptic conductances and dividing the input firing rates by the same factor (whereas
changing the level of background input corresponds to
scaling firing rates without changing unitary synaptic
conductances). The average synaptic conductance remains constant with this manipulation because it is proportional to the product of the unitary synaptic conductance and the input firing rate. On the other hand, the
input current variance increases because it is proportional to the square of the unitary synaptic conductance
times the input rate.
Figure 3B shows the effects of using the dynamic
clamp to increase input noise without changing the
mean levels of total excitatory or inhibitory conductance
(control, closed circles; with additional noise, open
squares). It is well known that input noise enhances
neuronal responses (see, for example, Hô and Destexhe,
2000), as seen in this figure. However, an important and
less appreciated feature shown in Figure 3B is that the
enhancement is not uniform across different levels of
Gain Modulation from Background Synaptic Input
777
Figure 3. Separate Effects of Shunting and
Noise on Response Gain and Variability
(A) Firing rate versus constant driving current
for a neuron without (closed circles), and with
(open squares) 32 nS of additional constant
conductance in the absence of any additional
noise from background synaptic input. The
result is a pure shift of the firing-rate curve.
(B) Firing rate versus constant driving current
for a different neuron in the 1X condition
(closed circles) and with the same level of
conductance but input noise equivalent to the
3X condition (open squares). The effect is an
increase in firing rate that is largest at low
rates, resulting in a change in the slope of
the firing-rate curve.
(C) The firing-rate curves for the 1X (closed
circles) and 3X (open squares) conditions replotted from Figure 2A for comparison with
(A) and (B).
(D) The coefficient of variation (CV) of the interspike intervals for the low (closed circles)
and high (open squares) noise conditions shown in (B), plotted against firing rate. Increasing the level of input noise increases the CV.
(E) The CV of the interspike intervals for the 1X (closed circles) and 3X (open squares) conditions shown in (C), plotted against firing rate.
Increasing the level of background synaptic input has little effect on the CV.
(F) Standard deviation of the membrane potential (open diamonds) in the absence of any driving current, for which the neuron in (C) did not
fire, and the CV of the interspike intervals (closed diamonds) when the same neuron was driven to fire at approximately 20 Hz, plotted as a
function of the rate factor. Both measures of response variability are roughly constant and independent of the rate factor, other than when it
is zero.
driving current or, equivalently, firing rates. The rateenhancing effect of input noise is most pronounced at
low firing rates, and it decreases steadily as the firing
rate increases. This induces a slope change in the firingrate curve (Figure 3B; n ⫽ 6). In other words, increasing
input noise increases the magnitude of the neuronal
response for a given level of driving current, but it decreases the sensitivity of the response to changes in
this current. Only the first of these effects has been
widely discussed (although see Doiron et al., 2001).
Although a complete analysis of the effects of noise on
neuronal firing is complex (see, for example, Ricciardi,
1977) we can provide an intuitive explanation for why
noise enhances firing rates and why this effect diminishes at high rates. In the absence of noise and at a low
firing rate, the membrane potential spends an appreciable amount of time hovering slightly below the threshold
for action potential generation before the neuron fires.
Noise-induced fluctuations can cause the neuron to fire
earlier than it would have fired in the absence of noise
at any time during this hovering period. This effect
causes the increase in firing rate at low rates. However,
when the neuron is firing rapidly, the membrane potential rises quickly to threshold without a significant hovering period. In this situation, balanced excitation and
inhibition have roughly equal and opposite average effects on spike timing. Thus, at high firing rates, noise
produces little net enhancement of firing rate.
Gain modulation by background synaptic input, as
seen in Figure 3C, is the combined result of a noiseinduced slope change (Figure 3B) and a shuntinginduced rightward shift (Figure 3A). A consequence of
these combined effects is that, unlike what happens
with noise alone (Figure 3B), increasing the level of background input decreases response gain without the accompanying overall enhancement of firing (Figure 3C,
1X, closed circles; 3X, open squares). In other words,
increasing balanced synaptic input decreases the sensitivity of the response to changes in driving current while
decreasing the magnitude of the neuronal response for
a given level of driving current. In summary, an increase
in input noise causes a leftward shift and a slope change
in the firing-rate curve, and shunting causes a rightward
shift. A balanced increase in background synaptic input
causes both to occur, and the left and right shift cancel,
leaving a pure gain decrease.
Effects of Background Input
on Response Variability
Although varying the amount of input noise is crucial
for changes in gain, gain modulation by this mechanism
is not accompanied by significant changes in the response variability of a neuron. We quantified the degree
of response variability by measuring the coefficient of
variation (CV) of the interspike intervals and the variance
of the subthreshold membrane fluctuations for the recorded neurons. Although increasing the amount of input noise with a fixed level of shunting (as in Figure
3B) increases response variability (Figure 3D, control,
closed circles; additional noise, open squares), changing the level of background synaptic input (as in Figure
3C) does not (Figure 3E, 1X, closed circles; 3X, open
squares; and Figure 3F). When background input rates
were quadrupled from the 1X condition, the CV did not
change appreciably (Figures 3E and 3F, closed diamonds), and the standard deviation of subthreshold
membrane potential fluctuations changed by less than
30% (Figure 3F, open diamonds; these were measured
at a level of driving current that produced no action
potentials). Thus, background synaptic input acts as a
source, but not as a modulator, of response variability.
Neuron
778
Instead, as we have shown, it acts as a modulator of
response gain.
Response variability does not vary appreciably with
the level of background synaptic input over the range
we consider because the effects of the increased input
noise are cancelled by increased synaptic shunting. We
can illustrate this directly in a passive neuron model
(without spiking) for the case of subthreshold membrane
potential fluctuations (also see Figure 3F for experimental data). The synaptic current arising from incoming
Poisson spike trains that activate conductances with an
instantaneous rise time and a decay time constant ␶s
has a correlation function that decays exponentially with
time constant ␶s, and which has an amplitude equal to
␴I2 , the variance of the synaptic current. In a passive
neuron model, the membrane potential is simply a lowpass filtered version of the synaptic current, and the
relationship between the current and voltage fluctuations can be computed in a straightforward manner to
give ␴V2 ⫽ ␴I2␶s/(g2 (␶m ⫹ ␶s)) for the variance of the voltage
fluctuations, where g is the total conductance (the sum
of the intrinsic membrane conductance and the average
synaptic conductance) and ␶m is the effective membrane
time constant. For the parameters we use, the total conductance of the neuron is dominated by synaptic contributions (typically the background synaptic conductance
is two or more times bigger than the membrane conductance) and ␶m ⬎ ␶s. Under these conditions, when balanced synaptic input rates are scaled by a rate factor
x, ␴I2 → x␴I2 , g → xg, and ␶m → ␶m/x to a good approximation. This makes the variance of the fluctuations in the
membrane potential approximately independent of the
rate factor, because ␴V2 ⬇ ␴I2␶s/(g2␶m) → x␴I2␶s/(x2g2␶m/
x) ⫽ ␴V2. Increasing background synaptic input rates
increases both the variance of the input current and
the overall conductance in such a way that there is no
significant increase in the variance of membrane potential fluctuations. This translates into approximately constant spike-train variability as well.
Gain Modulation in a Model Neuron
Gain modulation through background synaptic input
(Figure 3C) relies on a combination of the change in the
slope of the firing-rate curve due to input noise (Figure
3B) and the rightward shift of the curve due to shunting
(Figure 3A). A multiplicative effect, as in Figure 2E, will
arise only if the rightward shift of the curve is of the
appropriate magnitude (for example, Figure 4D shows
what happens when the shift is too small). Furthermore,
achieving gain modulation over an appreciable range of
firing rates requires sufficiently high levels of input noise.
To explore the sensitivity of gain modulation to these
requirements, we used a modeling approach in which
parameters can be varied freely.
We observed gain modulation through background
synaptic input in a variety of models, ranging from relatively realistic conductance-based descriptions to simplified integrate-and-fire neurons. Here, we report results using a particularly convenient approach, an
analytic expression for the firing rate of an integrateand-fire neuron receiving noisy input (Ricciardi, 1977).
To adapt this approach for our purposes, we approximate the background synaptic input by an equivalent
Figure 4. Gain Modulation in a Model Neuron
(A) Firing rate versus driving current from the analytic model for the
1X (thick trace), 2X (dashed trace), and 4X (thin trace) conditions.
The standard deviation of the membrane potential fluctuations was
5 mV.
(B) Gain versus firing-rate curves for 1X (thick trace), 2X (dashed
trace), and 4X (thin trace) conditions as in (A).
(C) Same as (B) but for membrane potential fluctuations of 2 mV.
(D) The thick trace is the firing rate versus driving current in the 1X
condition described in (A). For the thin trace, the excitatory input
rates were scaled by 5.3 and the inhibitory rates by 4.
white-noise current source and an equivalent shunting
conductance (see Appendix). The resulting expression
for the firing rate depends on the driving current I, the
total conductance of the neuron g, and the variance of
the membrane potential fluctuations produced by the
white noise, ␴V2. Because ␴V2 is approximately independent of the level of background input, as seen in Figure
3F and discussed above, we simplify the analysis by
treating it as a constant.
The analytic model displays gain modulation analogous to that shown in Figures 2A and 2B for the recorded
neurons (Figures 4A and 4B, 1X, thick trace; 2X, dashed
trace; 4X, thin trace). Although the shapes of the firing
rate and gain curves are different for the model and real
neurons, noise affects the real and model neurons in
qualitatively similar ways. We found that the real neurons
were more sensitive to noise than the model neurons
we studied, and therefore gain modulation effects could
be obtained in the real neurons using lower levels of
input noise than in the model.
The rate-dependent enhancement of firing due to input noise (Figure 3B) fades at high firing rates. Because
gain modulation relies on this effect, the gain changes
are restricted to firing rates below some critical value.
For the mechanism to be a viable candidate for the types
of gain modulation seen in vivo, this critical rate must be
sufficiently high. Figure 4B shows the gain modulation
produced by changing background input rates when the
noise-induced membrane potential fluctuations have a
standard deviation of 5 mV. In this condition, the model
Gain Modulation from Background Synaptic Input
779
reveals significant gain modulation over the entire range
of firing rates from 0 to 200 Hz. However, when the
membrane potential fluctuations have a standard deviation of only 2 mV, gain modulation diminishes above
about 50 Hz (Figure 4C, 1X, thick trace; 2X, dashed
trace; 4X, thin trace). This indicates that noise levels of
approximately 5 mV are needed to make the mechanism
viable. We suggest that the large membrane potential
fluctuations measured in vivo (see, for example, Anderson et al., 2000a) may be present to support gain modulation through background synaptic input over a wide
range of firing rates.
Doiron et al. (2001) have noted previously, in a modeling study of voltage-dependent inhibition in neurons of
the electrosensory lateral line lobe, that the combined
effects of shunting and noise can have a multiplicative
effect on firing rates. However, the effect they reported
was limited to low firing rates because it involved noise
arising solely from inhibitory input. This produces considerably smaller voltage fluctuations than the balanced
combination of excitation and inhibition that we study.
For gain modulation to be multiplicative, as in Figure
2E, the amount of shunting and the amount of noise
introduced by the background synaptic input must be
appropriately matched. This might be a source of concern because, in our experiments, the dynamic clamp
simulates synaptic conductances located at or near the
soma of the neuron. Distal synapses can contribute to
the input noise, but they may not produce much
shunting. In the balanced configuration we use, shunting
arises primarily from inhibitory inputs, and the noise is
dominated by excitation. Thus, our results should apply
even if the excitatory inputs are too distal to produce
shunting, provided that inhibitory synapses are proximal
enough to do so.
However, another effect can compensate if the
shunting-induced shift of the firing-rate curve is too
small (or, for that matter, if it is too large). Recall from
Figure 1A, that ordinary excitation and inhibition also
shift the firing-rate curve. Therefore, any deficiencies in
the amount of shifting produced by shunting can be
compensated for by adjusting the ratio of excitation to
inhibition that make up the background synaptic input.
For example, introducing a slight excess of inhibition
over excitation within the mixture of background synaptic input can compensate for the fact that distal inputs
might produce insufficient shunting to completely shift
the firing-rate curve to its appropriate position, as we
have verified in modeling studies.
Nevertheless, multiplicative gain modulation requires
an appropriate combination of the effects of shunting
and the degree of balance between excitation and inhibition. Figure 4D shows what happens if this condition is
not met. In this example, the excitatory rate was increased 1.3 times as much as the inhibitory rate when
background input was increased (rather than in a oneto-one ratio as for all the other figures). This has no
effect on the noise-induced slope change of the firingrate curve, but it causes the rightward shift seen in the
transition from Figure 3B to 3C to be incomplete, so
that the firing-rate curves cross. Thus, instead of the
firing-rate curves for different gains converging at zero,
they converge at approximately 35 Hz in this example.
Although the curves in Figure 4D no longer look like
Figure 5. Modulating the Gain of the Response to an Oscillatory
Input
The firing rate of an integrate-and-fire model neuron driven by sinusoidally oscillating (1 Hz) current. For times less than 2.5 s, the neuron
was in the 1X condition (thick trace in Figure 4D). At 2.5 s, the
background input to the neuron was increased as it was for the thin
trace in Figure 4D, but the oscillating input remained the same.
The resulting gain modulation reduced the amplitude, but not the
average level, of the response.
the gain-modulated curves in Figure 4A, they still display
a form of response scaling that can be a useful feature,
as illustrated in Figure 5. In this example, an oscillating
driving current caused oscillations in the firing rate of an
integrate-and-fire neuron. When the level of background
synaptic input was modified at 2.5 s during the middle
of the trace in Figure 5, the amplitude of the response
oscillations, but not the average level of the response,
decreased. This result, which is quite distinct from what
could be obtained by conventional inhibitory effects,
arises because the crossing point for the firing-rate
curves was equal to the average response rate.
While the effect shown in Figure 5 is fairly modest,
significantly larger effects can be obtained when neurons are connected together in a network. Recurrent
excitation in such a circuit can amplify neuronal responses. In this case, reducing the gain of the neurons
has two effects: it reduces their responsiveness to external input; and it also decreases the impact of the excitatory drive that they exert on each other. Because of
this dual effect, the impact of gain modulation on the
responses of network neurons can be considerably
larger than what is seen for the single neuron in Figure
5, as we have verified in model networks (results not
shown).
We can use the analytic model to investigate one
more issue related to gain modulation, the nature of the
multiplicative scaling produced by background synaptic
input (Figure 2E). Examination of the analytic expression
for the firing rate indicates that the relevant scaling factor in the model is g␴V (see Appendix). Over the range
we study, this factor varies in proportion to the rate
factor x. In the Appendix, we show that, under appropriate conditions, the dependence of the firing rate on
the rate factor x and the driving current I is of the form
xF(I/x), where F is a nonlinear function. Thus, multiplicative gain modulation of the firing rate in the analytic
model arises from a combination of divisive scaling of
the current and multiplicative scaling of the firing rate.
Although gain modulation in more realistic models and
in real cells may exhibit more complex scaling laws, this
Neuron
780
example provides a tractable case illustrating the basic
features of the effect.
Discussion
Predictions
The mechanism we propose predicts that gain modulating signals, whether proprioceptive (Andersen and
Mountcastle, 1983; Andersen et al., 1985), attentional
(McAdams and Maunsell, 1999a; Treue and Martı́nezTrujillo, 1999), or from other sources (Heeger, 1992; Peña
and Konishi, 2001), are carried by a combination of excitatory and inhibitory inputs with firing rates that rise and
fall together. Gain modulation by background synaptic
input has two critical signature features. First, increases
in gain are associated with decreases in background
synaptic input and neuronal conductance. Second, the
change in conductance during gain modulation is not
accompanied by a significant modification in the variance of the membrane potential fluctuations or the variability of the spiking response. This last prediction
agrees with data showing that spike-train variability was
not affected by shifts in attention that multiplicatively
scaled the amplitude of neural responses (McAdams
and Maunsell, 1999b).
The background input that controls gain may arise
from many sources: local circuits, distal sources, or
feedback pathways (Bastian, 1986; Koch and Ullman,
1985; Hupé et al., 1998; Przybyszewski et al., 2000).
If local excitatory and inhibitory synaptic connections
provide significant, balanced synaptic input, a strong
population response evoked by a potent stimulus (such
as a high-contrast image in the case of the visual system)
will generate a high level of balanced synaptic input that
will lower response gain. This can produce the type of
divisive response normalization that has been proposed
to account for a number of observed phenomena in
primary visual cortex, including response saturation
(Heeger, 1992).
Background input from distal sources could be responsible for effects associated with attention, by increasing gain for attended stimuli. Interactions between
locally and distally generated background inputs could
explain why little attentional gain modulation is seen
for high-contrast visual images (Reynolds et al., 2000).
According to the proposed mechanism, attentional enhancement requires a reduction in the background input
from sources controlling attention. However, for highcontrast images, the large amount of locally generated
background input responsible for response saturation
could mask the attention-related reductions in background input coming from distal sources.
Finally, gain modulation could be generated by background synaptic input arising from feedback pathways.
Although such pathways are excitatory and tend to terminate on distal dendrites, they could generate balancing inhibition and the accompanying conductance
changes by exciting local interneurons. If so, strong
responses in a secondary area, which is the source of
the feedback signal, could reduce the gain of responses
in a primary sensory area. Conversely, neurons in the
primary area would remain at high gain if they failed
to generate a strong response in the secondary area
controlling their gain. In addition to assuring that signals
of different efficacy all get transmitted from the primary
to the secondary area, such feedback can lead to popout effects in which responses to unexpected stimuli,
such as visual patches that differ from a background
pattern, are enhanced (Rao and Ballard, 1999).
Conclusion
Sherman and Guillery (1998) have proposed that neurons have two classes of inputs, one responsible for
driving neural responses and the other for modulating
those responses. At the level of ionotropic synapses
that could generate both fast responses and rapid modulation, it is not clear what anatomical basis exists for
this dual classification. We suggest that the two classes
of inputs are not defined anatomically, but rather, functionally. Sets of balanced inputs that have excitatory
and inhibitory rates rising and falling together comprise
modulatory inputs, and those for which excitation and
inhibition vary in opposite directions act as driving inputs. This arrangement has the advantage that individual excitatory inputs can rapidly switch between driving
and modulatory functions, depending on whether they
are varying in parallel with or in opposition to changes
in inhibition.
Neurons receiving large amounts of background excitatory and inhibitory input operate quite differently
from the conventional picture of neuronal input integration. Our observation that levels of background input
modify response gain suggests that such activity is not
simply a source of response variability, but instead,
plays an important role in controlling cortical processing. In summary, we propose that cortical neurons
in vivo operate in a transistor-like mode in which sets
of excitatory and inhibitory inputs with covarying firing
rates act as gain control signals to gate other sets of
driving inputs with opposing, push-pull excitation and
inhibition.
Experimental Procedures
Slices of rat somatosensory cortex were prepared for whole-cell
recording as described previously (Reyes and Sakmann, 1999). Rats
were anesthetized with halothane and decapitated. One hemisphere
of the brain was then removed, and 300 ␮m slices were cut in icecold artificial cerebrospinal fluid or ACSF (125 mM NaCl, 25 mM
NaHCO3, 25 mM glucose, 2.5 mM KCl, 1.25 mM NaH2PO4, 2 mM
CaCl2, 1 mM MgCl2), using a vibratome tissue slicer. During recording, slices were perfused with 30⬚C ACSF bubbled with 95%
O2, 5% CO2. Simultaneous whole-cell somatic recordings from layer
5 pyramidal neurons were performed under dynamic clamp using
two electrodes, patched at the soma, filled with 100 mM K-gluconate, 20 mM KCl, 10 mM phosphocreatine, 10 mM HEPES, 4 mM
ATP-Mg, and 0.3 mM GTP at pH 7.3 (filled electrode resistances
ranged from 5 to 10 M⍀). The dynamic clamp is a voltage-controlled
current clamp (Sharp et al., 1993; Robinson and Kawai, 1993)
that uses an analog multiplier to calculate and inject the current
that would be produced by computer-determined conductance
changes. With the dynamic clamp, one electrode was used for current injection and the other to record the neuronal membrane potential. Recordings were terminated if neurons were spontaneously
firing or had an initial resting potential above ⫺55 mV, or if either
electrode had an initial access resistance greater than 50 M⍀.
Excitatory and inhibitory synaptic currents were calculated using
an analog multiplier as Isyn ⫽ gsyn (Erev ⫺ V ), where gsyn is the computercontrolled synaptic conductance generated from simulated presynaptic spike trains, Erev is the reversal potential of the synaptic con-
Gain Modulation from Background Synaptic Input
781
ductance (0 mV for excitatory inputs and ⫺80 mV for inhibitory
inputs), and V is the measured membrane potential of the neuron.
Presynaptic spike trains were generated by Poisson processes at
the specified rates. The unitary synaptic conductance for each presynaptic spike was calculated as a difference of exponentials, with
time constants of 0.1 ms for the rising phase and either 5 ms (excitatory) or 10 ms (inhibitory) for the falling phase. The peak unitary
synaptic conductances were set to 2% (excitatory) or 6% (inhibitory)
of the measured resting membrane conductance. The same parameters were used for all the neurons studied.
A step of constant driving current lasting 2–3 s was injected along
with the simulated background synaptic input. The firing rate was
measured after 500 ms of stimulation by counting the number of
action potentials over the remaining stimulus time interval and dividing by its duration. Trials were separated by a recovery period of
at least 10 s.
The analytic formula used for Figure 4 describes the firing rate of
an integrate-and-fire model with input consisting of white noise and
constant current and is discussed in the Appendix. In addition to
the parameters given in the text, the firing rate in the analytic model
depends on the resting value of the membrane time constant (taken
to be 20 ms) and the difference between the action potential threshold and reset membrane potentials (taken to be 6 mV following
the analysis of Troyer and Miller [1997]). For the 1X condition, we
assumed that the background synaptic input produced a synaptic
conductance equal to the resting conductance of the neuron.
The integrate-and-fire model in Figure 5 had a resting membrane
potential Vrest ⫽ ⫺65 mV, an action potential threshold Vth ⫽ ⫺54
mV, and a postspike reset potential Vreset ⫽ ⫺60 mV. The synaptic
dynamics was identical to that used for the real neurons, except
that the rise of the synaptic conductances was taken to be instantaneous. The parameters describing the background synaptic input
were adjusted to produce subthreshold membrane potential fluctuations with a standard deviation of 5 mV (total excitatory and inhibitory
synaptic conductances equal to 0.4 and 1.6 times the resting membrane conductance, with rates of 135 Hz for both). The firing rate in
response to an oscillating driving current was extracted by counting
spikes and averaging over multiple runs.
Appendix
The analytic expression for the firing rate r of an integrate-and-fire
neuron (with resting membrane conductance g0, synaptic conductance gs, resting membrane time constant ␶0, and action-potential
threshold and reset voltages Vth and Vreset), receiving a constant
current I and white noise current that produces membrane potential
fluctuations with standard deviation ␴V, takes the form
r⫽
冢
冣
g␴v
V ⫺ 0.5(Vth ⫹ Vreset)
f ∞
,
g0(Vth ⫺ Vreset)␶0
␴v
where g ⫽ g0 ⫹ gs is the total membrane conductance and V∞ ⫽
(g0V0 ⫹ gsVB ⫹ I)/g. V0 is the resting potential of the neuron and VB is
the potential at which the excitatory and inhibitory synaptic currents
sum to zero. An explicit expression for the function f, which is an
integral of a combination of Gaussian and error functions, is not
needed for our discussion (Ricciardi, 1977), and an additional dependence on the quantity (Vth ⫺ Vreset)/␴V has been suppressed because
this is a constant in the case we consider. To simplify the analysis,
we take V0 ⫽ VB ⫽ 0.5(Vth ⫹ Vreset) from which it follows that V∞ ⫽
0.5(Vth ⫹ Vreset) ⫹ I/g and
r⫽
冢 冣
g␴v
I
f
.
g0(Vth ⫺ Vreset)␶0 g␴v
As discussed in the text, the combination g␴V varies approximately
in proportion to the scale factor x when background input rates are
varied. Thus, the above equation implies the relationship given in
the text, r ⬀ xF(I/x), with F closely related to f . The function f initially
rises from zero with a nonlinear dependence on its argument, but
for sufficiently large values of I it becomes linear. When this occurs,
the factor of g␴V within f and the factor of g␴V multiplying f in the
above equation cancel, and gain modulation goes away, as discussed in the text.
Acknowledgments
This research was supported by the Sloan-Swartz Foundation, NEIT32-EY-7136, NSF-IBN-9817194, and NSF-IBN-0079619.
Received: March 19, 2002
Revised: June 19, 2002
References
Andersen, R.A., and Mountcastle, V.B. (1983). The influence of the
angle of gaze upon the excitability of light-sensitive neurons of the
posterior parietal cortex. J. Neurosci. 3, 532–548.
Andersen, R.A., Essick, G.K., and Siegel, R.M. (1985). Encoding of
spatial location by posterior parietal neurons. Science 230, 450–458.
Anderson, J.S., Lampl, I., Gillespie, D.C., and Ferster, D. (2000a).
The contribution of noise to contrast invariance of orientation tuning
in cat visual cortex. Science 290, 1968–1972.
Anderson, J.S., Carandini, M., and Ferster, D. (2000b). Orientation
tuning of input conductance, excitation, and inhibition in cat primary
visual cortex. J. Neurophysiol. 84, 909–926.
Bastian, J. (1986). Gain control in the electrosensory system mediated by descending inputs to the electrosensory lateral line lobe.
J. Neurosci. 6, 553–562.
Bernander, Ö., Douglas, R.J., Martin, K.A.C., and Koch, C. (1991).
Synaptic background activity influences spatiotemporal integration
in single pyramidal cells. Proc. Natl. Acad. Sci. USA 88, 11569–
11573.
Borg-Graham, L.J., Monier, C., and Frégnac, Y. (1998). Visual input
evoles transient and strong shunting inhibition in visual cortical neurons. Nature 393, 369–372.
Dayan, P., and Abbott, L.F. (2001). Theoretical Neuroscience (Cambridge, MA: MIT Press), p. 164.
Destexhe, A., and Paré, D. (1999). Impact of network activity on the
integrative properties of neocortical pyramidal neurons in vivo. J.
Neurophysiol. 81, 1531–1547.
Destexhe, A., Rudolph, M., Fellous, J.-M., and Sejnowski, T.J. (2001).
Fluctuating synaptic conductances recreate in vivo-like activity in
neocortical neurons. Neuroscience 107, 13–24.
Doiron, B., Longtin, A., Berman, N., and Maler, L. (2001). Subtractive
and divisive inhibition: effect of voltage-dependent inhibitory conductances and noise. Neural Comput. 13, 227–248.
Hahnloser, R.H., Sarpeshkar, R., Mahowald, M.A., Douglas, R.J.,
and Seung, H.S. (2000). Digital selection and analogue amplification
coexist in a cortex-inspired silicon circuit. Nature 405, 947–951.
Hansel, D., and Van Vreeswijk, C. (2002). How noise contributes
to contrast invariance of orientation tuning in cat visual cortex. J.
Neurosci. 22, 5118–5128.
Heeger, D.J. (1992). Normalization of cell responses in cat striate
cortex. Vis. Neurosci. 9, 181–198.
Hirsch, J.A., Alonso, J.-M., Reid, R.C., and Martinez, L.M. (1998).
Synaptic integration in striate cortical simple cells. J. Neurosci. 18,
9517–9528.
Hô, N., and Destexhe, A. (2000). Synaptic background activity enhances the responsiveness of neocortical pyramidal neurons. J.
Neurophysiol. 84, 1488–1496.
Holt, G.R., and Koch, C. (1997). Shunting inhibition does not have
a divisive effect on firing rates. Neural Comput. 9, 1001–1013.
Holt, G.R., Softky, W.R., Koch, C., and Douglas, R.J. (1996). Comparison of discharge variability in vitro and in vivo in cat visual cortex
neurons. J. Neurophysiol. 75, 1806–1814.
Hupé, J.M., James, A.C., Payne, B.R., Lomber, S.G., Girard, P.,
and Bullier, J. (1998). Cortical feedback improves discrimination
between figure and background by V1, V2 and V3 neurons. Nature
394, 784–787.
Koch, C., and Poggio, T. (1992). Multiplying with synapses and neurons. In Single Neuron Computation, T. McKeena, J. Davis, and S.
Zornetzer, eds. (Orlando, FL: Academic Press), pp. 315–345.
Neuron
782
Koch, C., and Ullman, S. (1985). Shifts in selective visual attention:
towards the underlying neural circuitry. Hum. Neurobiol. 4, 219–227.
McAdams, C.J., and Maunsell, J.H.R. (1999a). Effects of attention
on orientation-tuning functions of single neurons in macaque cortical area V4. J. Neurosci. 19, 431–441.
McAdams, C.J., and Maunsell, J.H.R. (1999b). Effects of attention
on the reliability of individual neurons in monkey visual cortex. Neuron 23, 765–773.
Mel, B.W. (1993). Synaptic integration in an excitable dendritic tree.
J. Neurophysiol. 70, 1086–1101.
Miller, K.D., and Troyer, T.W. (2002). Neural noise can explain expansive, power-law nonlinearities in neural response functions. J. Neurophysiol. 87, 653–659.
Peña, J.L., and Konishi, M. (2001). Auditory spatial receptive fields
created by multiplication. Science 292, 249–252.
Pouget, A., and Sejnowski, T.J. (1997). Spatial transformations in the
parietal cortex using basis functions. J. Cogn. Neurosci. 9, 222–237.
Przybyszewski, A.W., Gaska, J.P., Foote, W., and Pollen, D.A. (2000).
Striate cortex increases contrast gain of macaque LGN neurons.
Vis. Neurosci. 17, 485–494.
Rao, R.R., and Ballard, D.H. (1999). Predictive coding in the visual
cortex: a functional interpretation of some extra-classical receptivefield effects. Nat. Neurosci. 2, 79–87.
Rapp, M., Yarom, Y., and Segev, I. (1992). The impact of parallel
fiber background activity on the cable properties of cerebellar Purkinje cells. Neural Comput. 4, 518–532.
Reyes, A., and Sakmann, B. (1999). Developmental switch in the
short-term modification of unitary EPSPs evoked in layer 2/3 and
layer 5 pyramidal neurons of rat neocortex. J. Neurosci. 19, 3827–
3835.
Reynolds, J.H., Pasternak, T., and Desimone, R. (2000). Attention
increases sensitivity of V4 neurons. Neuron 26, 703–714.
Ricciardi, L.M. (1977). Diffusion Processes and Related Topics in
Biology (Berlin: Springer-Verlag).
Robinson, H.P.C., and Kawai, N. (1993). Injection of digitally synthesized synaptic conductance transients to measure the integrative
properties of neurons. J. Neurosci. Methods 49, 157–165.
Salinas, E., and Abbott, L.F. (1995). Transfer of coded information
from sensory to motor networks. J. Neurosci. 15, 6461–6474.
Salinas, E., and Abbott, L.F. (1996). A model of multiplicative neural
responses in parietal cortex. Proc. Natl. Acad. Sci. USA 93, 11956–
11961.
Salinas, E., and Abbott, L.F. (1997) Invariant visual responses from
attentional gain fields. (1997). J. Neurophysiol. 77, 3267–3272.
Salinas, E., and Thier, P. (2000). Gain modulation: a major computation principle of the central nervous system. Neuron 27, 15–21.
Shadlen, M.N., and Newsome, W.T. (1994). Noise, neural codes and
cortical organization. Curr. Opin. Neurobiol. 4, 569–579.
Sharp, A., O’Neil, M.B., Abbott, L.F., and Marder, E. (1993). The
dynamic clamp: artificial conductances in biological neurons.
Trends Neurosci. 16, 389–394.
Sherman, S.M., and Guillery, R.W. (1998). On the actions that one
nerve cell can have on another: distinguishing “drivers” from “modulators”. Proc. Natl. Acad. Sci. USA 95, 7121–7126.
Smith, M.R., Nelson, A.B., and du Lac, S. (2002). Regulation of firing
response gain by calcium-dependent mechanisms in vestibular nucleus neurons. J. Neurophysiol. 87, 2031–2042.
Softky, W.R., and Koch, C. (1993). The highly irregular firing of cortical cells is inconsistent with temporal integration of random EPSPs.
J. Neurosci. 13, 334–350.
Srinivasan, M.V., and Bernard, G.D. (1976). A proposed mechanism
for the multiplication of neural signals. Biol. Cybern. 21, 227–236.
Stevens C.F., and Zador, A.M. (1998). Input synchrony and the irregular firing of cortical neurons. Nat. Neurosci. 3, 210–217.
Tiesinga, P.H.E., José, J.V., and Sejnowski, T.J. (2001). Comparison
of current-driven and conductance-driven neocortical model neurons with Hodgkin-Huxley voltage-gated channels. Phys. Rev. E.
Stat. Phys. Plasmas Fluids Relat. Interdiscip. Topics 62, 8413–8419.
Treue, S., and Martı́nez-Trujillo, J.C. (1999). Feature-based attention
influences motion processing gain in macaque visual cortex. Nature
399, 575–579.
Troyer, T.W., and Miller, K.D. (1997). Physiological gain leads to high
ISI variability in a simple model of a cortical regular spiking cell.
Neural Comput. 9, 971–983.
Zipser, D., and Andersen, R.A. (1988). A back-propagation programmed network that simulates response properties of a subset
of posterior parietal neurons. Nature 331, 679–684.
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement