Skip to main content
  • Loading metrics

Mechanisms underlying a thalamocortical transformation during active tactile sensation


During active somatosensation, neural signals expected from movement of the sensors are suppressed in the cortex, whereas information related to touch is enhanced. This tactile suppression underlies low-noise encoding of relevant tactile features and the brain’s ability to make fine tactile discriminations. Layer (L) 4 excitatory neurons in the barrel cortex, the major target of the somatosensory thalamus (VPM), respond to touch, but have low spike rates and low sensitivity to the movement of whiskers. Most neurons in VPM respond to touch and also show an increase in spike rate with whisker movement. Therefore, signals related to self-movement are suppressed in L4. Fast-spiking (FS) interneurons in L4 show similar dynamics to VPM neurons. Stimulation of halorhodopsin in FS interneurons causes a reduction in FS neuron activity and an increase in L4 excitatory neuron activity. This decrease of activity of L4 FS neurons contradicts the "paradoxical effect" predicted in networks stabilized by inhibition and in strongly-coupled networks. To explain these observations, we constructed a model of the L4 circuit, with connectivity constrained by in vitro measurements. The model explores the various synaptic conductance strengths for which L4 FS neurons actively suppress baseline and movement-related activity in layer 4 excitatory neurons. Feedforward inhibition, in concert with recurrent intracortical circuitry, produces tactile suppression. Synaptic delays in feedforward inhibition allow transmission of temporally brief volleys of activity associated with touch. Our model provides a mechanistic explanation of a behavior-related computation implemented by the thalamocortical circuit.

Author summary

We study how information is transformed between connected brain areas: the thalamus, the gateway to the cortex, and layer 4 (L4) in cortex, which is the first station to process sensory input from the thalamus. When mice perform an active object localization task with their whiskers, thalamic neurons and inhibitory fast-spiking (FS) interneurons in L4 encode whisker movement and touch, whereas L4 excitatory neurons respond almost exclusively to touch. To explain these observations, we constructed a computational model based on measured circuit parameters. The model reveals that without touch, when thalamic activity varies slowly, strong inhibition from FS neurons prevents activity in L4 excitatory neurons. Brief and strong touch-induced thalamic activity excites both excitatory and FS neurons in L4. FS neurons inhibit excitatory neurons with a delay of approximately 1 ms relative to ascending excitation, allowing L4 excitatory neurons to spike. Our results demonstrate that cortical circuits exploit synaptic delays for fast computations. Similar mechanisms likely also operate for rapid stimuli in the visual and auditory systems.


Thalamocortical circuits represent model systems for multi-area computations [1]. Sensory information enters the cortex through the thalamus. Transformations in thalamocortical circuits have mostly been studied in anesthetized animals with passive sensory stimuli [28] or with artificial whisking [9]. In the somatosensory system these studies have revealed subtle differences in receptive field structure across neurons in the thalamocortical circuit [24].

However, active sensation in awake animals involves dynamic interactions with the world, such as saccades [10], palpation with the digits of the hand [11], or movements of the whiskers on the face of rodents [1214]. During active sensation, movement of the sensors produces ‘reafferent’ signals, whereas interactions with the world generate ‘exafferent’ signals. During haptic exploration, movement activates peripheral sensors to produce reafference and touch generates exafference [1521]. The brain needs to parse these different signals for perception [13]. During active sensation, movement attenuates the transmission of certain sensory signals to the cortex [2224]. Tactile suppression is thought to enhance perception of salient events that cannot be predicted based on movement. Tactile suppression is an example of adaptive filtering [25, 26], which is critical for low-noise encoding of relevant sensory stimuli. Here we identify the mechanisms of adaptive filtering in the thalamocortical circuit of the mouse whisker system.

Whisker touch and movement are transduced by mechanosensory afferents in the whisker follicle. Information then flows through the trigeminal ganglion, to the brainstem, thalamus (barreloids in the ventral posterior medial thalamic nucleus, VPM) and terminates in the primary somatosensory cortex (vS1). The main target of VPM axons is the Layer 4 (L4) barrels in vS1. The microcircuit of each L4 barrel is mostly contained within the barrel, and the connections between specific cell types within the barrel have been mapped: L4 excitatory neurons and L4 fast-spiking, parvalbumin (PV)-expressing GABAergic interneurons (FS) are connected within type and across types [27, 28]. Apart from neuromodulation, the only known long-range input to L4 originates in VPM [2931]. VPM excites all L4 neuron types, and L4 FS neurons inhibit the excitatory neurons to implement feedforward inhibition [3235].

Cell type-specific recordings from VPM, L4 excitatory [36] and L4 FS neurons [37] uncovered a fundamental computation performed by L4 barrels. Neurons in VPM respond to touch, but they also increase their activity during whisker movement (‘whisking’) [1721]. L4 FS neurons have nearly identical dynamics as VPM neurons. In contrast, L4 excitatory neuron spikes are strongly coupled to touch, but respond only weakly to whisking [36]. L4 microcircuits therefore transmit touch signals and suppress reafferent signals generated by whisking. These observations indicate that the thalamocortical circuit accentuates salient tactile information by suppressing signals related to self-movement.

Multiple observations regarding the L4 circuit remain to be explained [37]. First, the circuit suppresses self-movement signals but transmit touch signals. From a theoretical perspective, this selective filtering violates the linear response expected from theories of strongly coupled cortical networks [38, 39]. Second, whereas the baseline spike rates of VPM and L4 FS neurons are substantial during baseline conditions and more than double during whisking, the spike rate of L4 excitatory neurons is very low at baseline and does not increase during whisking. Third, when an inhibitory opsin (halorhodopsin) is photostimulated in L4 FS neurons, the spike rates of FS neurons decrease and the spike rates of L4 excitatory neurons increase [37]. Although this result is naively expected, it contradicts the "paradoxical effect" predicted from models of neural circuits that are stabilized by inhibition and from models of strongly-coupled networks [4042]. According to these models, both inhibitory and excitatory neurons should respond with increased spike rates at modest levels of inactivation of inhibitory neurons [41].

To understand the mechanisms underlying these experimental observations we constructed and analyzed a conductance-based model of the L4 circuit. The synaptic circuits of L4 of the barrel cortex have been studied in great detail in brain slices [2729, 4346]. These measurements allowed us to constrain the numbers of neurons, neuronal properties, the patterns of connectivity, and the average synaptic strengths in the model. We set parameter values near values extracted from these measurements. This "reference set of parameters" was adjusted over a restricted range, such that the model circuit displayed dynamics similar to the actual L4 circuit. We then varied parameters to explore their roles in controlling system dynamics. We studied how the model network responds to thalamic input at baseline, during whisker movement and during touch. We used the model to disentangle the roles of feedforward synaptic connections from the thalamus and recurrent intracortical connections in shaping L4 dynamics. The model revealed an important role for synaptic delays, fast synaptic kinetics, and inhibitory and excitatory conductance strengths in shaping the L4 responses during behavior.


Thalamocortical transformation of tactile information

We first summarize recordings made in VPM and in L4 from excitatory and FS GABAergic interneurons in mice performing an object location discrimination behavior (Fig 1a and 1b). Head-restrained mice had to localize a vertical pole with their whiskers for a water reward [47, 48]. Whisker movement and touch were tracked on millisecond time scales with high-speed videography [49, 50] (Fig 1a). Recordings were made with extracellular silicon probe recordings in VPM [37], and loose-seal cell-attached and whole cell recordings in L4 [36, 37, 51].

Fig 1. Summary of experimental results in the object-localization task [37].

(a) Top, schematic illustrates measurement of whisker position (azimuthal angle θ), instances of touch and an example trace of whisker position. Protraction corresponds to positive changes in θ. (b) Schematics of the thalamocortical circuit and relevant cell-types. (c) Spike rate aligned to transitions from non-whisking to whisking (adapted from panel 5e in [37]). (d) Average spike rate as a function of whisking amplitude. (e) Average population response aligned to touch (adapted from panel 5c in [37]). Data and figures corresponding to previously reported datasets [36, 37].

VPM neurons have significant spike rates (mean, 5.1 Hz), even in the absence of whisker movement and touch (Table 1; Fig 1c and 1d; see Figs 2 and 3 for representative examples) [37]. VPM neurons were also highly sensitive to whisking (Fig 2b and 2c) [17, 18, 21]. Spike rates increased after whisking onset (average, 3-fold; Fig 1c and 1d). Given that mice whisk at approximately 15 Hz under these conditions [36], the spike rates during whisking correspond to approximately one spike per whisking cycle, on average. The modulation depths with whisking amplitude and phase (Supplementary Figure 2 in [37]) were similar to published studies in rats [18]. VPM neurons also responded to active touch with a brief increase in spike rate (0.6 spikes per touch) (Fig 1e). The peri-stimulus time histogram (PSTH) aligned to touch onset shows a sharp peak in spike rate, with short latency after touch (3.1±0.6 ms) and brief duration (2.9±1.7 ms) (Fig 2i, 2n and 2s). Exafferent touch signals and reafferent whisking signals were multiplexed in individual VPM neurons (Fig 2).

Table 1. Summary experimental results for VPM, L4 excitatory and L4 FS neurons.

Data are taken from [37]. The increase of the average spike rate of L4 excitatory neurons from non-whisking to whisking conditions was found to be insignificant. Values are reported as mean ± standard deviation. In comparison to VPM neurons, the average response of L4 excitatory neurons is reduced by a factor of 2 in response to touch and by a factor of 20–30 in response to whisking onset.

Fig 2. Example neurons recorded in VPM during whisker-based object localization.

(a) Chronic silicon probe recordings in VPM. (b,c) Two example trials from the same neuron showing increases in activity with whisker movements without touch. Green line, whisker azimuthal angle. Black ticks, spikes. (d,e) Two example trials showing increases in spike rate after touch. (f) Touch rasters (black dots; sorted by last touch in a trial). The magenta line shows when the pole was moved within reach of the whiskers. The green line represents the last touch in each trial. (g) Whisker movement amplitude. Red rectangle, epoch of high whisker movement amplitude and high spike rate (same as in h). Orange rectangle, epoch of low amplitude and low spike rate (same as in h). (h) Spike raster for an example neuron in VPM (blue dots, spikes). The black arrow in g-h indicates onset of whisker twitching [52]. The twitching is triggered by an auditory signal generated by the brief activation of a shutter. (i) Peri-stimulus-time-histogram (PSTH) showing response for touch. (j) Spike rate as a function of whisker movement amplitude. (k-o) Same as f-j for another example VPM neuron with low baseline spike rate. (p-t) Same as f-j for another example VPM neuron that does not show modulation with whisking amplitude.

Fig 3. Example neurons across the thalamocortical circuit.

(a) Touch raster (black dots; sorted by last touch in a trial). The magenta line shows when the pole was moved within reach of the whiskers. The green line represents the last touch in each trial. (b) Whisker movement amplitude. Red rectangle, epoch of high whisker movement amplitude and high spike rate. Orange rectangle, epoch of low amplitude and low spike rate. (c) Spike raster of an example neuron in VPM (blue dots, spikes). Red and orange regions of interest correspond to B. (d) PSTH showing response for touch. (e) Spike rate as a function of whisker movement amplitude. (f-j) Same as a-e for a L4 FS neuron. (k-o) Same as a-f for a L4 excitatory neuron. The few spikes that occur during whisking are phase-locked to movement.

The modulation of L4 FS neurons to touch and whisker movement were similar to VPM neurons (Figs 1c–1e, 2 and 3). L4 FS neurons increased their spike rate after onset of whisking (average, 3-fold). L4 FS neurons responded reliably to touch. The response onset had slightly longer latency (5–15 ms) and longer duration (2–15 ms) than for VPM neurons. The increased latency is expected from the propagation time delay between VPM and L4 (2 ms) [37, 53].

The responses of L4 excitatory neurons differed profoundly from those of VPM and L4 FS neurons [36] (Figs 1c–1e and 3). The baseline spike rate of L4 excitatory neurons was much lower and did not increase significantly after onset of whisking. The response after touch onset occurred with longer latency and longer duration than for the VPM neurons, and similar to L4 fast-spiking neurons (Fig 1e).

The transformation performed by L4 circuits can be summarized by three main findings (Fig 1c–1e; Table 1). First, baseline spike rates are high in VPM and L4 FS neurons, and low in L4 excitatory neurons. Second, VPM and L4 FS neurons elevate their spike rates further during periods of whisker movement, whereas L4 excitatory neuron spike rates remain low. The spike rates of VPM and L4 FS neurons during periods of whisking are more than one order of magnitude larger than those of L4 excitatory neurons. Third, all three neuron types respond to touch reliably (Table 1; Fig 1e). The touch responses of the three neuronal populations are the integrals under the curve in Fig 1e. The average touch response of L4 excitatory neurons (0.3 spikes/touch) is about half of the average touch response of VPM neurons, whereas the spike rates of L4 excitatory neurons are lower than those of VPM neurons by a factor of 20–30. The VPM response is more transient than the L4E response. The L4 circuit therefore performs a behaviorally relevant computation by propagating information related to external stimuli and suppressing responses to whisker movement, a predictable stimulus.

The dynamics of FS neurons during whisker movement suggest that these neurons suppress whisker movement-related activity in L4 excitatory cells. Consistent with this view, optogenetic reduction of activity in a subset of L4 FS neurons expressing the light-gated, inhibitory chloride pump eNphHR3.0 (L4I-Hr+) unmasks movement-related activity in L4 excitatory neurons [37] (Fig 4a and 4b). Photostimulation of L4I-Hr+ neurons decreases the activity in L4 FS neurons on average and increases the activity in L4 excitatory neurons. The suppression of L4I-Hr+ neurons causes an increase in response in L4 excitatory neurons during whisker movements (Fig 4c) and an overall increase in the number of spikes per touch (Fig 4d). This result implies that suppression of whisking response in L4 excitatory neurons involves inhibition from FS neurons. It remains unclear, however, why touch responses in L4 excitatory neurons are not suppressed as well.

Fig 4. Summary of effect of photoinhibition of L4 FS neurons on L4 neurons.

(a) Schematic of the Hr+ experiment. A subset of L4 FS neurons do not express Hr (Hr- neurons). (b) Photostimulation of Hr+ neurons decreases the activity in L4 FS neurons and increases the activity in L4 excitatory neurons. Black: without photostimulation; orange: with photostimulation. (c) Response of L4 excitatory neurons during whisker movements (adapted from panel 8d in [37]). (d) Response to touch of L4 excitatory neurons. The black circle denotes the mean values (adapted from panel 8e in [37]).

Membrane potential measurements with whole-cell recordings provide additional clues about mechanisms [37]. L4 excitatory neurons depolarize substantially (6 mV) and briefly after touch. Touch-related inhibitory input (from FS neurons) to L4 excitatory neurons is delayed by approximately 0.5 ms with respect to excitatory input (Fig 1e). It has been proposed that this short ‘window of opportunity’ [2, 35] allows L4 excitatory neurons to spike after touch, typically with one spike [36], before inhibition suppresses L4 excitatory activity. During whisker movement, excitation is matched by inhibitory input, keeping the L4 excitatory neuron membrane potential well below spike threshold. Suppression of self-movement signals is therefore implemented by inhibition within L4.

Yu et al. [37] also showed that activating halorhodopsin in L4I-Hr+ neurons during whisker movement suppresses activity in these neurons. This result poses a new question. The major cellular effect of activation of halorhodopsin is hyperpolarization [54, 55]. Theoretical investigations of cortical circuits, consisting of recurrently connected excitatory neurons stabilized by inhibitory neurons, have investigated the effects of hyperpolarization of inhibitory neurons [41]. Over a large range of conditions, hyperpolarization of inhibitory neurons leads to increased spike rates in both excitatory and also inhibitory neurons. This "paradoxical effect" is common to multiple network regimes, including networks with strong synaptic conductances that fire in an asynchronous manner [40] and networks with moderate synaptic conductances, if the excitatory-to-excitatory synaptic coupling gEE is sufficiently strong [41]. Does the L4 circuit lie outside of the modeled parameter regimes, or can other factors explain the lack of paradoxical effect?

A computational model of tactile suppression in L4

To explain the dynamical response of the L4 circuits to thalamic input, and to understand the roles of feedforward and recurrent connections, we constructed and analyzed a detailed computational model of the L4 circuit. L4 excitatory and L4 FS neurons make connections within type and across types with high connection probability (Fig 5a; Table 2). The thalamocortical circuit of the rodent whisker system is one of the most extensively studied mammalian circuits. In vitro and in vivo studies have measured many fundamental parameters that are required to construct a realistic computational model of the thalamocortical circuit. Our model circuit consists of 1600 L4 excitatory (E) neurons and 150 L4 FS (I) neurons [28] receiving input from 200 VPM (T) neurons (Fig 5a). When referring to modeling results we denote VPM neurons as T, L4 excitatory neurons as L4E, and GABAergic interneurons as L4I (Fig 5b and 5c). Since we know little about non-FS GABAergic interneurons during behavior we model only one inhibitory neuronal population.

Fig 5. Neural network model of L4.

(a) Diagram of the recurrent model of L4 network. (b) Spike shape of VPM (schematic), L4E and L4I neurons. (c) Temporal dynamics of individual EPSPs for the different synaptic connections (T = VPM; I = L4 FS; E = L4E). The convention is that that the first letter corresponds to the post-synaptic neuron and the second letter to the presynaptic neuron. (d) Thalamic generating function FT (Eq 1). The panels on the right show the same figure in a magnified scale. For simplicity, we assume that all T neurons have the same preferred phase.

Table 2. Network parameters in the reference parameter set: - synaptic delay, Kαβ—average number of presynaptic inputs, gαβ—synaptic conductance, Vextr—the extremal value of the unitary synaptic membrane potential change.

The corresponding experimental values for Kαβ and Vextr are written in the two right columns. Those values are taken from the following references: a—[53], b—[56], c—[34], d- [28], e–[57], f—[27].

Cortical neurons were simulated using a conductance-based model with a single compartment per neuron [58, 59]. Synaptic conductances are denoted as gαβ, where β is the presynaptic population and α is the postsynaptic population (Fig 5a). Model parameters, including numbers of neurons, unitary synaptic conductance, connection probability, resting potential, and membrane time constants are based on neurophysiological measurements in brain slices [27, 28, 56, 60] with small adjustments (see Methods). VPM provides the only external input to the model circuit, with connectivity estimates based on in vivo and brain slice measurements [34, 53]. Spike trains of T neurons were modeled as independent inhomogeneous Poisson processes with a generating function FT representing thalamic activity during quiescence, whisking or whisking and touch (Fig 5d; see Methods). (1) where AT is the spike rate averaged over a whisker movement cycle, BT is the modulation depth, and CT is the number of spikes per touch, τw is the whisking period, tc is the time of touch onset within a whisking cycle, n is the cycle number, and Θ is the Heaviside function. During whisker movements AT increases above a baseline, with sinusoidal modulation phase-locked to a single preferred phase ϕ. Touch is represented by adding a rectangular function at touch onset t = tc (with respect to the whisking cycle), stretched over τc = 3 ms with an integral of CT = 0.6 spikes per touch (Eq 1). Correlations between thalamic neurons beyond those generated by Eq 1 have not been measured and were thus not modeled. The population-average spike rate of neurons within a population over whisking cycles is denoted by να (α = T,E,I), and the population-average spiking response to touch is denoted by Rα. Note that without touch, νT = AT. A ‘reference parameter set’ is described in Methods and used unless otherwise stated.

Response of the model circuit to whisker movement and touch

We start by simulating the model for whisking without touch (thalamic spikes per touch, CT = 0). L4E spike rates were low (νE < 1 Hz) compared to those of L4I neurons (Fig 6a and 6b). The average rates νE and νI depend on AT (the population- and time-average thalamic spike rate during whisking only) but are almost independent of the modulation depth BT (Eq 1). Except near threshold, νI was proportional to AT, in consistent with theoretical results for large and sparse neuronal networks with strong synapses, in which strong excitation is compensated by inhibition (balanced networks) [38]. The L4E spike rate νE increased linearly with AT, but with a much smaller slope compared to νI.

Fig 6. A neural network model of L4 explains suppression of whisker movement signals in L4 excitatory neurons.

The colors black, grey and red denote T, L4E and L4I neuronal populations respectively. (a) Example L4E and L4I membrane potential during simulated whisking (green). (b) The population- and time average spike rates νE and νI of the L4E and L4I neurons respectively as function of the thalamic input AT in the absence of touch. L4I neurons follow linearly the thalamic input while L4E neurons increase only weakly with AT beyond firing threshold. Inset, zoom in. (c) Membrane potential for an example neuron during whisking and touch (black dots). (d) Population PSTH aligned to touch onset. Inset, zoom in.

The linearity of the νI-AT curve fit empirical observations (Table 1, Fig 1c; spike rates of both VPM and L4 FS neurons more than double in the transition from non-whisking to whisking). The average spike rate of L4 excitatory neurons, νE, is low [37]. During transition from non-whisking to whisking, the average spike rate of L4 excitatory neurons barely changes (from 0.4 Hz to 0.6 Hz; Table 1). The spike rates of L4E neurons in our model show similar dynamics.

L4E neurons fire at most one spike each after touch, whereas L4I neurons fire more than one (up to two) spikes per touch (Fig 6c and 6d). The synaptic delay, , is a critical factor in determining the strength of the touch response (Fig 7a). For , the average normalized response of L4E neurons after touch, RE, is small (RE = 0.01 spikes/touch), and L4I neurons fire RI = 0.64 spikes/touch. Strong inhibition overwhelms excitation before L4E neurons have a chance to spike. The delay between the appearance of excitation and feed-forward inhibition in the L4E neuron has been termed ‘window of opportunity’ [2, 35, 6163]. This ‘window of opportunity’ increases with , but is not equal to it because it is affected by the durations of thalamocortical synaptic process and the time needed for the L4I to fire in response to the brief and strong thalamic input. For a more realistic value, , L4E and L4I fire 0.34±0.24 and 1.3±0.07 spikes/touch respectively, similar to experimental measurements (Table 1). The touch responses for L4E and L4I (Fig 6d) are narrower than those seen in real data (Fig 1e). The detailed shapes of the modeled responses depend on network parameters (e.g., see Fig 8d–8i below). In addition, heterogeneity among neurons, which was not modeled here, is expected to broaden the responses.

Fig 7. Touch response in function of synaptic delay, AMPA receptors’ time constant and parameters defining thalamic input.

In each panel, responses to touch in L4E and L4I neurons (RE and RI, in grey and red respectively) are plotted, as well as the thalamic response in black. Spikes per touch were counted up to 25 ms after touch onset, and baseline computed by counting spikes 25 ms before touch is subtracted. Responses to touch are plotted as functions of (a) I-to-E synaptic delay , (b) the AMPA receptor time constant tAMPA, (c) the thalamic response to touch, CT, and (d) the thalamic spike rate AT during whisker movements without touch.

Fig 8. Effects of varying intracortical recurrent excitatory conductances gEE on the function of L4E neurons.

(a) The circuit with changing gEE emphasized in green. (b) νE vs. AT during whisker movements only, for 11 values of gEE from 0 (light green) to 0.4 mS/cm2, that is twice the reference parameter value (dark green). Recurrent excitation gEE increases νE while not affecting the slope of the νE-AT curve far from spiking threshold substantially. (c) RT, RE and RI as functions of gEE. Other parameters: AT = 14 spikes/s, CT = 0.6. (d) PSTH aligned to touch onset for L4E and without recurrent excitation (gEE = 0 mS/cm2). (e) Same as C for L4I. (f-g) Same as c-d for gEE = 0.2 mS/cm2. (h-i): Same as c-d for gEE = 0.35 mS/cm2. Beyond ~gEE = 0.4 mS/cm2 the network exhibits runaway excitation.

The response of the network to touch depends on the time-course of synaptic conductances. Without touch signals, L4E neurons are inhibited by L4I neurons and spike at low rates (Fig 6a and 6b). Touch produces strong, brief and synchronous thalamic excitation (Fig 1e), which depolarizes L4E neurons and enables them to fire before inhibition terminates the response. This mechanism demands that excitatory synaptic conductance changes at thalamocortical synapses are brief. Touch responses of L4E and L4I neurons decrease with tAMPA, the decay time of AMPA-mediated EPSCs (Fig 7b). Substantial touch responses in L4E neurons require tAMPA < 2–3 ms, consistent with the brief excitatory conductances measured at thalamocortical synapses [60].

Touch responses of L4E and L4I neurons increased with the strength of the thalamic touch signal (Fig 7c), consistent with graded responses to touch strength measured in L4E neurons [36]. The response saturates at one and two spikes/touch for L4E and L4I respectively; this is in part due to the intrinsic properties of these neurons, which preclude them from firing more spikes in response to brief thalamic input.

We have shown that a L4 network with parameters similar to experimentally determined values can replicate major experimental findings. L4E excitatory neurons exhibit low baseline activity, low activity in response whisker movement, and significant response to touch. However, during behavior the responses of L4 excitatory neurons is not all-or-none: L4 excitatory neurons are tuned to multiple sensory features including touch direction, intercontact interval, strength of touch, and likely other factors [36]. Our simulation results indicate that L4E and L4I neurons respond to touch in a graded manner and robustly respond to touch at different whisking amplitudes. Specifically, our simulations show that RE and RI decrease with AT (Fig 7d) for all parameter values consistent with the data. Mechanistically, increasing thalamic input causes larger inhibition in the cortical circuit that decreases touch response. Future experiments could test this model prediction.

Contributions of intracortical connections in tactile suppression

The model allows us to explore how specific synaptic connections within L4 contribute to fine-tuning L4 function. These connections currently cannot be specifically manipulated, making them difficult to evaluate experimentally. Recurrent excitation (gEE) can only be tuned over a limited range before runaway excitation is triggered (for about gEE~0.4 mS/cm2), even during baseline or whisking (Fig 8a and 8b). The conductance gEE does not modify the slope of the νE-AT curve (Fig 8b), but shifts this curve upward. gEE amplifies touch responses RE and RI (Fig 8c) [64], mostly by increasing the duration of touch responses (Fig 8d–8i).

The dynamics of L4E neurons are shaped in subtle ways by L4 inhibition. Fig 9 shows how parameters involving inhibition (gII, gEI, gIE) affect circuit responses, while holding the other parameters at their reference values. Intracortical excitation of inhibition (gIE) is necessary to prevent runaway excitation, and thus keeps the spike rates of neurons in the network moderate (Fig 9a and 9b). Reducing gIE shifts the νE-AT curve during whisking upward, without modifying its slope. For touch responses, gIE shifts the touch-response RE-CT curve to the right while decreasing the slope of the linear section of this sigmoid (Fig 9c).

Fig 9. Inhibition in L4 controls the response to L4E neurons.

(a-c) Changing gIE from 0 to 1.2 mS/cm2 (d-f) Changing gEI from 0 to 1.4 mS/cm2. (g-i) Changing gII from 0 to 1.1 mS/cm2. In the top panels, the synaptic connection that its strength is varied is plotted in green. In the middle and bottom panels, curves are plotted for 11 values of the relevant g from 0 (light green) to its maximal value, that is twice the reference parameter value (dark green). (b, e, h) The response of L4E neurons, νE, to slowly varying thalamic input, which correlates with the amplitude of whisking. (c, f, i) The response of L4E neurons to spikes, RE, associated with touch.

Inhibition to L4E neurons (gEI) is also required to prevent runaway excitation (Fig 9d and 9e). For low values of gEI (but above values for runaway excitation) and large values of gII (Fig 9g and 9h), νE scales linearly with thalamic input AT. This linear scaling is known from balanced networks with strong synaptic coupling [38, 40]. The situation is different for sufficiently strong inhibition (relatively high gEI, moderate gII), where the response of L4E neurons to slowly-varying thalamic input during whisking, νE, is independent, and even decreasing, with AT (Fig 9e and 9h). For small enough gII, L4E are quiescent except of near firing threshold. In addition, low gET and large gIT values are critical to keep νE low for all whisker movement amplitudes (Fig 10a, 10b, 10d and 10e). Furthermore, these low values νE are nearly independent of AT. This regime is consistent with experimental data, where L4 FS neurons increase their spike rate with whisker movement, whereas L4 excitatory neuron spike rates remain low (Fig 1d, Table 1).

Fig 10. Effect of varying thalamocortical conductances on the function of L4E neurons.

Symblols and lines are as in Fig 9. (a) Changing gET. (b) νE vs. AT during whisker movements only. (c) RE vs. CT. (d) Changing gIT. (e-f) Same as b-c.

Increasing values of gEI and gII shifts the touch-response RE-CT curve to the right and to the left respectively (Fig 9f and 9i). In addition, gEI, but not gII, decreases the slope of the linear section of this sigmoid (Fig 9f). Similarly, gET and gIT shift the RE-CT curve to the left and to the right respectively (Fig 10c and 10f).

The effects of halorhodopsin activation in L4 FS neurons

Yu et al. ([37], Fig 8, supplementary Fig. 13) expressed halorhodopsin in FS neurons in L4 and explored how the responses of excitatory and FS neurons in L4 under baseline conditions, whisking and touch vary under halorhodopsin activation. L4 excitatory neurons increase their spike rates and L4 FS neurons decrease their spike rates, both during baseline and whisking conditions. Similarly, L4 excitatory neurons increase their spiking responses to touch and L4 FS neurons decrease their responses.

The halorhodopsin expression levels in FS neurons likely varied widely across individual neurons [65]. We therefore divided the L4I neurons in the model into hr expressing neurons and non-expressing neurons (Hr+ and Hr- respectively). The fraction of Hr+ neurons among L4I neurons is denoted by fhalo.

Manipulation of the components of neural networks can cause complex and counterintuitive change in network dynamics (‘paradoxical’ response; [4042, 66]: injecting negative current to all inhibitory neurons in a network, that includes spiking excitatory neurons, increases the average spike rates in inhibitory neurons). If the synaptic conductance strengths are moderate and the excitatory-to-excitatory synaptic conductance is above a certain level, injecting negative current to the inhibitory neurons causes the spike rates of excitatory neurons νE to increase, and as a result the network dynamics causes the spike rates of inhibitory neurons νI to increase as well [41, 42, 66]. Alternatively, if synaptic conductances are strong and the excitatory population is active, the condition that the activity of excitatory neurons should be moderate (non-zero and not epileptic) causes νI to increase under negative current injection to inhibitory interneurons [40]. The major effect of halorodopsin is to hyperpolarize FS neurons (see Methods). Therefore, if all inhibitory neurons in the model are assumed to express halorodopsin equally (fhalo = 1), halorhodopsin activation will cause both L4E and L4I neurons to increase their spike rates.

We simulated the response of L4E and L4I neurons to whisking for fhalo = 1 without and with halorhodopsin activation, and replicated the 'paradoxical effect': the average responses of L4E and L4I neurons to whisking, νE and νI, increase with simulated light activation. The simulation result for L4I neurons is in contrast to experimental observations showing that most L4 FS neurons increase their spike rates. This discrepancy is resolved if we assume fhalo = 0.5. For this value, spike rates during whisking increase, on average, for L4E neurons (Fig 11a), decrease for almost all L4I-Hr+ neurons (Fig 11b), and increase for inhibitory neurons that do not express halorhodopsin (L4I-Hr-).

Fig 11. Simulated light activation of halorhodopsin expressed in L4I-Hr+ neurons.

Simulations with fhalo = 0.5 reveals a reduction in the whisking suppression and an enhancement of touch responses by L4E neurons. (a) Halorhodopsin activation in L4I-Hr+ causes an average increase in response of L4E during whisking and no touch, with a wide distribution of halorhodopsin—induced modifications. (b) Most L4I-Hr+ neurons reduce their activity during whisking while L4I-Hr- neurons increase it. (c) Increase in the touch responses in L4E neurons during suppression of L4I-Hr+. (d) Increase in the touch responses in L4I neurons. The increase in touch responses is only seen in Hr+ cells. (e,f) Population PSTH of L4E (e) and L4I (f) neurons with and without L4I-Hr+ activity suppression. (g) Reduction of L4I-Hr+ activity diminishes the whisking suppression effect in L4E neurons. Black line: T neurons; solid grey line: L4E neurons without halorhodopsin activation; dashed grey line: L4E neurons during halorhodopsin activation. (h) Reduction of L4I-Hr+ activity diminishes the whisking response in L4I-Hr+ neurons. Solid red line: L4I-Hr+ neurons without halorhodopsin activation; dashed red line: L4I-Hr+ neurons during halorhodopsin activation. Dashed blue line: L4I-Hr- neurons during halorhodopsin activation.

The spike responses of L4E neurons to touch increase moderately for fhalo = 0.5, especially for neurons with low baseline response (Fig 11c and 11e), whereas the spike response of L4I-Hr- inhibitory neurons remains about the same and that of L4I-Hr+ increases somewhat (Fig 11d and 11f). These dynamics stem from the fact that the initial response to touch, before feedforward inhibition hyperpolarizes the neuron, is determined mainly by feedforward excitation. While halorhodopsin activation hyperpolarizes inhibitory neurons and increase the difference between spiking threshold and their membrane potentials before touch onset, the driving force of excitation (the difference between the reversal potential of APMA-mediated excitation and their membrane potential) increases as well and enhances excitation.

The reduction in response in L4I-Hr+ neurons partially disrupts whisker movement suppression in L4E neurons, and increases the slope of the νE-AT curve (Fig 11g). During photostimulation of L4I-Hr+ neurons, their response to increasing whisker movements becomes shallower (Fig 11h). This change in response of L4-Hr+ neurons in turn causes a steeper response to whisker movements in L4E neurons (Fig 11g).

Finally we explored how the previous results depend on fhalo. The mean touch responses for all neuronal populations is non-monotonic: it increases when fhalo increases from zero, and then decreases for higher values (Fig 12a). For our parameter set, simulated halorhodopsin activation increases touch response of L4I-Hr+ neurons (in comparison with no activation) for fhalo < 0.78. During whisker movements, the spike rates of all neuronal populations increases with fhalo, and the average spike rate of L4I-Hr+ neurons during simulated halorhodopsin activation is lower than that with no activation for fhalo < 0.74 (Fig 12b). Interestingly, L4-Hr+ neurons reduce their spiking response to touch with halorhodopsin activation for large fhalo, but reduce their spike rates in response to whisking for small fhalo. For small fhalo, the spike rates of L4-Hr+ neurons in response to whisking are low (Fig 12b), because this small group of neurons is both suppressed by halorhodopsin activation and inhibited by the majority of L4-Hr- neurons. In contrast, the spiking responses to touch of L4-Hr+ and L4-Hr- neurons are similar and to the responses of L4I neurons without halorhodopsin activation. This behavior is obtained because touch responses are transient and are reduced by the global level of inhibition before touch. For low fhalo, population inhibition is similar to inhibition without halorhodopsin activation.

Fig 12. Effects of light activation of halorhodopsin expressed in L4I-Hr+ neurons.

Effects of halorhodopsin light activation on touch response and whisking response for the L4E (grey), L4I-Hr+ (red) and L4I-Hr- (blue) neuronal populations are plotted as functions of fhalo, the fraction of L4I-Hr+ neurons among all L4I neurons. In both panels, values without and with light activation are denoted by solid and dashed lines, respectively. (a) Spikes per touch, R, for L4 neuronal populations as functions of fhalo. (b) Average spike rates, ν, during whisking for L4 neurons as functions of fhalo.


We computationally dissected how information is transformed between thalamus and barrel cortex at the level of defined neural circuits [36, 37] (Fig 1). Our model defines the mechanisms that enable the L4 circuit to transmit touch-related information and suppress self-motion signals in L4. Here we summarize our main modeling results and how they explain major experimental observations.

Transmission of touch signals while suppressing whisking signals

We compared our model to recordings from VPM projection neurons (Figs 13), L4 FS neurons, and L4 excitatory neurons [37] during performance of an object localization task, while whisker movements and touches were tracked with millisecond time scale precision [36, 48, 51, 67]. VPM and L4 FS neurons respond to whisker movement and touch, whereas L4 excitatory neurons responded almost exclusively to touch. During whisking, excitation to L4 excitatory neurons from VPM is only slowly modulated in time and is matched by feedforward and feedback inhibition from L4 FS neurons, which cancels self-movement signals. In contrast to the self-movement input, touch-related inputs are brief and synchronous (Fig 1e). Our model establishes the conditions for a ‘window of opportunity’, in which L4 excitatory neurons can fire before inhibition catches up [35, 62, 64, 68, 69]. The I-to-E synaptic delay must be sufficiently large (say, ~1 ms) and the AMPA-mediated synaptic conductances should be brief (Fig 8a and 8b). The brief duration of this window diminishes the chance of L4 excitatory neurons to fire multiple spikes upon touch, producing low trial-to-trial variability in spike count after touch [36, 70].

Spike rates during baseline and whisking

During baseline and whisking, L4 FS neurons spike on average at tens of Hz, while L4 excitatory neurons spike on average below 1Hz. The average spike rates of VPM and L4 FS excitatory neurons, νT and νI, more than double after whisking onset, while the average spike rates of L4 excitatory neurons remains below 1 Hz. Low spike rates of L4E neurons require particular combinations of synaptic parameters: gIE, gEI and gIT need to be strong, and gEE, gII and gET should be weak (Figs 8c, 9, 10). If gET is too small or gIT is too large, L4E neurons will be quiescent. Similarly, L4E neurons will not spike if gII is too small (Fig 9h). For moderate gII, νE can remain about constant with AT.

Touch responses

The response of L4E neurons to touch, RE, increases with CT (the thalamic response to touch) in a sigmoid manner, because the intrinsic properties of L4E neurons in the model impose a refractory period that in general precludes rapid firing of more than one spike. Synaptic parameters that increase νE during whisking shift the sigmoid function leftward, and those that decrease νE shift that function rightward (Figs 9 and 10). In our model, the inhibitory neuronal population spikes at significant rates (20–40 Hz) even at baseline. This allows the number of spikes/touch RE to be smaller than one and vary gradually with input strength (Fig 9f). As a result, stronger thalamic responses to touch, for example as a result of stronger touch, generates proportionally stronger L4E responses, consistent with experimental results (Fig 1e) [36].

Effects of photostimulation of L4-Hr+ neurons

In response to halorhodopsin activation, L4 excitatory neurons increase their average spike rates and L4-Hr+ FS neurons reduce their spike rates [37]. The model does not replicate this behavior if halorhodopsin is "expressed" in all inhibitory neurons, because νI increases with simulated photostimulation. This discrepancy is resolved if a sufficient number of L4I neurons lack halorhodopsin (or express halorhodopsin at very low levels) (Figs 11 and 12b). With photostimulation, L4-Hr+ I neurons decrease their spike rates if their fraction among I neurons, fhalo, is below a certain value (0.74 in Fig 12a), while L4-Hr- FS neurons increase their spike rates. Future work is needed to determine whether including more populations of inhibitory interneurons, for example somatostatin-positive neurons [71] will generate parameter regimes in which L4-Hr+ neurons decrease their responses to both whisking and touch upon light activation.

Modeling approach

The development of powerful computing resources has enabled simulations of entire cortical columns of diverse neurons with realistic morphologies and biophysically plausible membrane conductances [72, 73]. Because of the profusion of parameters, detailed models are difficult to analyze to extract the underlying principles. So far these detailed models have failed to explain any neural computation [74]. Here we took a different approach. We implemented a computational model based on hard won numbers for connection probability, connection strength, neuron numbers and basic cellular parameters for the whisker thalamocortical circuit. The neurons themselves were generic single compartment models. Given the reduced nature of the model, it is possible to analyze the model in detail. This analysis yields testable predictions. For example, the dependence of average spike rates of neuronal populations on the average spike rates of thalamic neurons (Fig 6b). These predictions could be tested in experiments where stimuli, applied for example with a magnetic stimulator, are applied during whisker movement at different amplitudes.

The main characteristics of the thalamocortical circuit are: (i) Strong external inputs to fast-spiking inhibitory neurons and excitatory neurons. (ii) Strong inhibition within L4, implementing feedforward and lateral inhibition. (iii) Recurrent excitation. iv) A brief but nonzero delay between the spike of an inhibitory neuron and the onset time of the inhibitory post-synaptic conductance in a post-synaptic neuron. Parameters based from neurophysiological measurements produce a model circuit with behavior that is in qualitative agreement with in vivo measurements over a large range of conditions. However, it was necessary to adjust the model parameters slightly (i.e. within a factor of 2–3; Methods) to best match experimental data (reference parameter set, Methods). Experimental inaccuracy, sampling errors, and differences across biological conditions (such as in vitro brain slices versus behaving brain) likely preclude more accurate estimation of synaptic parameters. Changing even a single parameter could lead to quantitative, and sometimes even qualitative changes in behavior (Figs 810), as has been noted in other contexts [75]. This means that the detailed parameters matter. For example, if gII is too strong then L4I neurons do not respond sufficiently briskly to reduce whisking signals in L4E neurons (Fig 9g and 9h); on the other hand, if gII is too weak then inhibition shuts down L4E neurons. Similarly, gEE > 0 is required to amplify touch signals, but elevating gEE by a factor of two beyond the optimal level causes run-away excitation (Fig 8b). Moreover, the kinetics matter, such as the durations of the I-to-E synaptic delay (Fig 7a), axonal delays, the time-courses of synaptic conductances (Fig 7b), and intrinsic neuronal properties.

Comparison with previous L4 models

Rate models have been used to study cortical responses to fast-rising stimuli in the whisker system [8, 61] and elsewhere [71]. These previous models of L4 aimed to account for fast cortical responses to passive whisker deflections [8], but rate models cannot reliably describe brief responses to rapidly-varying stimuli [76]. Moreover, in important aspects these models have opposite behavior to our experimental and modeling results and to the known anatomy and physiology of L4 circuits. First, the modeled L4 neurons show activity in the absence of input, in contradiction to recent measurements [37, 77]. Second, L4 neurons exhibit strong and brief response to touch even for , in contrast to our model.

The propagation of synchronous and brief activity in cortical circuits has been investigated [2, 63, 64, 78, 79]. In these models, excitatory inputs produce strong inhibition in local circuits, which is slightly delayed with respect to the excitation. In L4 of the barrel cortex, the convergence of thalamocortical input onto L4 FS neurons is higher than for L4 excitatory neurons (probabilities of connections are 0.75 vs 0.4) [34] and thalamic stimuli produce larger synaptic potentials in L4 FS than L4 excitatory neurons [35]. These features allow propagation of brief synchronous activity across network layers. These models can respond to brief, strong stimuli by brief responses of the L4 E and I neuronal populations. Our model goes beyond this effect by showing how this brief touch response can be obtained together with large νI and small νE in response to baseline and whisking. For example, simple feedforward models such as [35] cannot exhibit small νE for a wide range of AT, corresponding to both baseline and whisking. That model, however, explores the development of cortical responses of L4E and L4I neurons over the long time scale of short-term synaptic plasticity. Conducting such a study in a model with recurrent connections and comparing its outcome to in vivo measurements remains to be carried out.

Models of networks of strongly-coupled neurons compensated by inhibition have been studied extensively [3840, 58, 80]. If inhibition on the excitatory neurons is not too large, the population-average response of excitatory and inhibitory neuron (νE and νI) scales linearly with the external (here, thalamic) input AT [39, 81]. The slope of the νE vs. AT increases with gII and decreases with gIE. If inhibition is too strong, νE = 0 and νI scales linearly with AT. In L4 and our model network, the numbers of neurons and the strengths of synaptic conductances are not very large. As a result, when νE is small (νE ≲ 1 Hz), similar to measured values [36, 37], νE can (except of near firing threshold, namely the AT values for which cortical neurons start firing) increase weakly with AT, be independent of AT, or even decrease with AT (Figs 9e, 9h, 10b and 10e). Similar behavior was obtained in rate models of V1 [82, 83]. This result mimics the empirical observation that during transition from non-whisking to whisking, the spike rates of VPM neurons more than double, the spike rates of L4I neurons increase proportionally, but the spike rates of L4E neurons remains low, with no significant increase with thalamic input strength. For larger νE (> 2 Hz), νE depends linearly on AT. Larger νE could be obtained with larger gII or smaller gEI (Fig 9b, 9e, 9h), or by suppressing a fraction of L4I neurons using halorhodopsin activation (Fig 11g).

Based on the linearity of strongly-coupled networks, one would expect that if responses of L4E neurons to whisking is an order of magnitude smaller than that of VPM neurons, the same proportions will maintain for the responses to touch, whereas experimentally the touch responses of L4 excitatory and VPM neurons differ by only a factor of two. We explain this effect by the fact that, with , inhibition lags excitation in response to brief and strong thalamic input. Indeed, for , the response to L4E neurons is significantly reduced (Fig 7a).

In this paper we emphasize the suppression of tactile reafference signals in the somatosensory cortex. However, reafference signals could still contribute to cortical computation. Consistent with previous studies [16, 18, 84], we observed that spike rates were modulated by the phase of the whisking cycle [37]. Spike rates were larger for certain phases of the whisker motion, and individual neurons had different preferred phases. Neurons in VPM and L4 all showed modulation with whisking phase. Excitatory and inhibitory inputs to L4 excitatory neurons showed phase-tuning. Across the population, the phase-tuning of inhibition from L4 fast-spiking neurons is biased towards protractions. This suggests that the touch-evoked responses in L4 excitatory neurons may be modulated by whisking phase, as has been previously suggested [84]. This kind of phase-dependent modulation could play a role in localizing objects by whisker touch during active sensation.

Materials and methods

Experimental methods

The experimental methods were described in [37]. In brief, we performed chronic multi-electrode silicon probe recordings from VPM and cell-attached recordings from L4 excitatory and L4 fast spiking neurons. To search for VPM neurons we lightly anesthetized the mice (0.6–1.2% isoflurane) and stimulated individual whiskers during extracellular recordings. We mapped the principal whisker (PW) to assess whether the PW was one of the large whiskers that can be tracked reliably during behavior. We stimulated several individual whiskers around the PW with a piezoelectric stimulator at multiple frequencies (i.e. 5,10,20,40 Hz) and recorded the neural activity. Before and/or after each day of behavioral recording we confirmed that neurons respond to stimulation of the PW (we re-positioned the electrode drive daily). In addition, after placing the animal in the behavioral apparatus, we usually maintained the anesthesia for several minutes to check that neurons still responded to the PW (by manual stimulation and/or by contacting the whiskers with the pole). Animals performed the behavioral task shortly after anesthesia was withdrawn.

Statistical analyses.

Data is presented as mean ± s.d. or mean ± s.e.m. as noted. Statistical comparisons use the non-parametric, two-sided Wilcoxon rank sum test. All statistical analyses were performed in Matlab (Mathworks).

Behavior, videography and recordings.

In [37] we recorded from VPM neurons in mice trained to localize an object using their whiskers [48]. In each trial, during a sample epoch lasting a few seconds (1.1–2.5, mean 1.9 s), a pole appeared in one of multiple locations on the right side of the head (Fig 1a). Animals had to withhold licking when the pole was presented in the most anterior location (typically out of reach of the whiskers), or lick to obtain a reward when the pole was in the posterior locations. High-speed videography and automated whisker tracking quantified whisker movement (azimuthal angle, θ; whisking phase, ϕ), changes in curvature caused by the forces exerted by the pole on the whisker (change in curvature, Δκ)[36, 50, 85], and contact time, all with 1 millisecond temporal precision [49, 51]. Mice touched the pole multiple times (mean number of touches, 8.9±4.2 before reporting object location with licking (mean reaction time 416 ± 165 ms; mean ± s.d.). Recordings were made in trained mice (71±7% correct responses). To examine the relationship of VPM activity and tactile behavior we aligned spikes of individual neurons recorded in VPM with high-speed recordings of whisker movement and touch [36, 50, 51].

Neuronal network model of a layer 4 barrel

Comprehensive neuroanatomical and neurophysiological data sets are revealing the connectivity between defined cell types over multiple spatial scales [28, 43, 44, 46, 86, 87]. But links between neural representations, computation and detailed anatomy are rarely achieved [6, 88]. One challenge is a lack of knowledge about strengths and dynamics of synapses between specific cell types during behavior. Most of such studies were carried out in slices, and there are differences between intrinsic and synaptic properties measured in slices and in vivo and between different in vivo states. Therefore, our strategy is to set parameter values (synaptic conductances, connectivity) close to measured in vitro values. However, differences between in vitro and in vivo conditions and experimental errors (e.g. estimates of unitary synaptic strengths) currently preclude exact quantification of the synaptic and connectivity properties (e.g., [53]). We define a set of parameter values, named "reference parameter set" [89], close to measured values when known (Fig 5), allowing for adjustment over a restricted range (mostly within a factor of 2 relative to empirical values), such that the circuit displays dynamics similar to experimentally-observed behavior. The reference parameter set specified below is used unless otherwise stated. Then, we vary one or two parameters to explore the role of those parameters on the system dynamics (e.g., Figs 810).

Network architecture.

We modeled one barrel in layer 4 (L4) of the vibrissa primary somatosensory cortex (vS1). Our model circuit consists of L4 excitatory (E) neurons and L4 FS (I) neurons [28] receiving input from VPM (T) neurons (Fig 5a). When referring to modeling results we denote VPM neurons as T, L4 excitatory neurons as L4E, and GABAergic interneurons as L4I (Fig 5b and 5c). Non-FS interneurons have not been studied during behavior and are therefore not included in the model.

Dynamics of single cortical neurons.

Intrinsic neuronal properties of L4 neurons have not been fully characterized. Therefore, following previous work on cortical dynamics [58, 59], single neurons are governed by a modified Wang-Buzsáki model with one compartment. The membrane potential , where α = E,I and i = 1,…, Nα obeys (2) where C is the neuron capacitance, (3) is the leak current, the conductance of the leak current is gL = 0.05 mS/cm2 and 0.1 mS/cm2 for the excitatory and inhibitory neurons, respectively, and VL = -65 mV. The current is the total synaptic input into the neuron. The other ionic currents are: where m = αm(V)/(αm(V)+βm(V)), IKdr = gKdr n4(V−VK), and IKZ = gKZ z(V-VK). The kinetics of h and n are given by: (4)

The functions α(V) and β(V), for V in mV, are: (5) (6) (7) (8) (9) (10)

We set ϕ = 0.2 to produce wider spikes than in [59]. The gating variable z of the adaptation current IKZ is governed by the equation: (11) Where z(V) = 1/{1+exp[-0.7(V+30)]} and τz = 60 ms. The parameters of the model are: gNa = 100 mS/cm2, VNa = 55 mV, gKdr = 40 mS/cm2, VK = -90 mV, C = 1μF/cm2. Excitatory neurons have gKZ = 0.5 mS/cm2, whereas inhibitory neurons do not possess this current (gKZ = 0).

Network architecture.

The network consists of NE excitatory neurons and NI inhibitory neurons. The L4 neurons receive input from NT thalamic relay neurons in one barreloid. For the simulations, NE = 1600, NI = 150, and NT = 200 [28]. Each neuron, whether excitatory or inhibitory, receives excitation and inhibition from the excitatory and inhibitory cortical neuronal populations respectively, as well as excitation from VPM thalamic neurons. The probability that a neuron from the βth, pre-synaptic population forms a synapse on a neuron from the αth, post-synaptic population is Kαβ/Nβ [90]. Therefore, a neuron from the αth population receives, on average, Kαβ synaptic inputs from neurons in the βth population. We define the matrix to be 1 if the jth neuron from the βth population projects to the ith neuron from the αth population, and 0 otherwise.

Synaptic input.

We consider only fast synapses corresponding to AMPA and GABAA receptors. The total synaptic input a cell i from population α receives is the sum of the synaptic currents from all the pre-synaptic populations: (12)

The total conductance for β = T,E,I is: (13) where the normalization constant τsyn,all is set to be 1 ms. The value denotes a synaptic variable of the pre-synaptic neuron j from the βth population projecting to the αth population. Those variables for AMPA and GABAA conductances are: (14)

The pre-synaptic neuron fires at times , k is the spike index, tsyn is the time constant of synaptic decay, Θ is the Heaviside function, and is the time delay between the firing by a spike from a presynaptic neuron from the βth population and the onset of the post-synaptic conductance in the post-synaptic neuron from the αth population. Unless specified otherwise, we use time constants tsyn = tAMPA = 2 ms for AMPA-mediated synapses and tsyn = tGABAA = 3 ms for GABAA-mediated synapses. The reversal potentials are VAMPA = 0 and VGABAA = -85 mV. The values of gαβ, Kαβ and are based on [27, 28, 34, 53, 56, 60] (Table 2). We scale the conductance strength like (Eq 13) to allow comparison of our conductance values gαβ with previous publications [40, 58], but analyze the circuits only for Kαβ values of corresponding to those of mouse L4 barrel and do not investigate the circuit dynamics as values of Kαβ vary. Compared to literature values the Kαβ were reduced by a factor of 2 for the following reasons. First, this reduction counteracts strong synchrony in the cortical circuits. Second, this could have been achieved by modeling highly variable synaptic strengths [91]. Third, L2/3 connectivity has a long tail of strong connections [92]. Similar heterogeneity in connections may appear also in L4. The heterogeneity increases the effective sparseness of the synaptic connection matrix. Assuming lower values of Kαβ is like taking into account strong connections only, including those in the tail of the distribution. The effect of this manipulation on the total synaptic conductance a neuron receives is within the factor 2–3 of resemblance between the experimentally-measured values (in slice experiments) and the values used in the model (Table 2).

The experimental evolution of the response of VPM and L4 neurons to consecutive touch events has not reported yet. In this work, we do not address how the responses to whisking and touch vary over the long time scale of short-term plasticity. For these reasons, and to keep the model and its analysis simple, the synapses in the model do not have depression and facilitation properties.

Spike rates of thalamic neurons.

Spike trains of T neurons were modeled as statistically-independent inhomogeneous Poisson processes. All T neuron share the generating function FT mimicking thalamic activity during quiescence, whisking or whisking and touch (Fig 5d): (15)

The first term in Eq 1 represents the baseline and whisking modulation contribution to the instantaneous spike rate, and the second term represents spikes added by touch. The parameter AT is the spike rate averaged over a whisker movement cycle without touch, BT is the modulation depth, and CT is the average number of additional spikes per touch, τw is the whisking period, tc is the time of touch onset within a whisking cycle, and n is the cycle number. During whisker movements, AT increases above a baseline, with sinusoidal modulation phase-locked to a single preferred phase ϕ (several neurons in VPM exhibit phase-locking activity to the whisking cycle;[37]). Touch is represented by adding a rectangular function at touch onset t = tc (with respect to the whisking cycle), stretched over 3 ms with an integral of CT = 0.6 spikes per touch (Eqs 1 and 15). We use the parameters: BT = 0.25 (based on the data of Supplementary Fig. 2d in [37]), τc = 3 ms, T = 100 ms, ϕ = π/2. The parameter tc is set to be 0.5T because thalamic neurons reach their maximal spike rate during whisking at around maximal retraction [37]. The parameters AT and CT are 6 Hz and 0 during no-whisking states, 14 Hz and 0 during whisking states and 14 Hz and 0.6 during whisking-and-touch states. For simplicity, we take the same values of AT, BT and CT for all thalamic neurons. Because of the large convergence of thalamic inputs into L4E and L4I neurons, our results are not modified significantly if we allow those thalamic parameters to be taken from a uniform distribution with the same average values.

Population- and time-averaged quantities.

We define να to be the population- and time-averaged spike rates of all the neurons in the αth population over many whisking cycles. Touch response of a neuron is defined to be the difference between the number of spikes fired by that neuron during a time window of 25 ms after touch onset and the number of spikes fired by the neuron 25 ms before touch onset. We define Rα to be the population- and time-average of the spiking response to touch of all the neurons in the αth population over many touch events.

Simulations of halorhodopsin activation.

We model activation of halorhodopsin, a light-gated chloride pump, expressed in L4I FS neurons (Figs 11 and 12). Halorhodopsin expression levels likely varied widely across individual neurons [65]. We therefore assume that halorhodopsin is expressed in only a fraction fhalo of the inhibitory neurons, 0 ≤ fhalo ≤ 1. Activation by light of the ith FS neurons that express halorhodopsin was modeled by two processes. First, as a chloride pump, its activation has an effect of negative current injection to the neuron. The amplitude of this current in the ith neuron is Ihalo,i = Ihalo,0 + Ihalo,1xi, where xi is taken randomly from a uniform distribution between -1 and 1, Ihalo,0 = -2 μA/cm2, and Ihalo,1 = 1 μA/cm2. As a result, Ihalo,i varies between -3 μA/cm2 and -1 μA/cm2. Second, the GABAA reversal potential is depolarized [93, 94] by a value ΔVGABAA,i = βIhalo,i, where β = -4mVcm2/μA. The second process has only a small quantitative effect on the spiking responses of cortical neurons.

Numerical methods.

Simulations were performed using the fourth-order Runge-Kutta method with time step Δt = 0.05 ms. Simulations with smaller Δt reveal similar statistics of neuronal firing patterns, such as spike rates ν (averaged over many whisking cycles), touch responses R or levels of synchrony. Differences between individual voltage time courses, however, diverged over large integration time interval. This divergence is expected because of the chaotic nature of the sparse network dynamics [38, 39, 95]. Statistics were computed after removing a transient of 0.5 s, over 60 s for panels showing dynamical properties of one realization (Figs 6a, 6c, 6d, 8d–8i and 11a–11f) and over 5.5 s for panel showing dependence on parameters (Figs 6b, 7, 8b, 8c, 9, 10, 11g, 11h and 12) where each data point is computed by averaging over 10 network realizations.


We thank David Kleinfeld, Rasmus Petersen, Rony Azouz and David Hansel for helpful discussions, Amy Hu, Brian Barbarits, Sun Wei-Lung and Tim Harris for help with experiments.

Author Contributions

  1. Conceptualization: DAG KS DG.
  2. Data curation: DG DAG KS.
  3. Formal analysis: DAG KS DG.
  4. Funding acquisition: KS DG.
  5. Investigation: DAG JY SAH MST MRB DG.
  6. Methodology: DAG KS DG.
  7. Project administration: KS DG.
  8. Resources: KS DG.
  9. Software: DG DAG KS.
  10. Supervision: KS DG.
  11. Validation: DG DAG KS.
  12. Visualization: DAG KS DG.
  13. Writing – original draft: DAG KS DG.
  14. Writing – review & editing: DG DAG KS.


  1. 1. Hubel D.H. and Wiesel T.N., Receptive fields, binocular interaction and functional architecture in the cat's visual cortex. J. Physiol. (London), 1962. 160: p. 106–154. pmid:14449617
  2. 2. Simons D.J. and Carvell G.E., Thalamocortical response transformation in the rat vibrissa/barrel system. J Neurophysiol, 1989. 61(2): p. 311–30. pmid:2918357
  3. 3. Brecht M. and Sakmann B., Dynamic representation of whisker deflection by synaptic potentials in spiny stellate and pyramidal cells in the barrels and septa of layer 4 rat somatosensory cortex. J Physiol, 2002. 543(Pt 1): p. 49–70. pmid:12181281
  4. 4. Brecht M. and Sakmann B., Whisker maps of neuronal subclasses of the rat ventral posterior medial thalamus, identified by whole-cell voltage recording and morphological reconstruction. J Physiol, 2002. 538(Pt 2): p. 495–515. pmid:11790815
  5. 5. Stanley G.B., et al., Visual orientation and directional selectivity through thalamic synchrony. J Neurosci, 2012. 32(26): p. 9073–88. pmid:22745507
  6. 6. Reid R.C. and Alonso J.M., Specificity of monosynaptic connections from thalamus to visual cortex. Nature, 1995. 378(6554): p. 281–4. pmid:7477347
  7. 7. Sun W., et al., Thalamus provides layer 4 of primary visual cortex with orientation- and direction-tuned inputs. Nat Neurosci, 2016. 19(2): p. 308–15. pmid:26691829
  8. 8. Kwegyir-Afful E.E., Kyriazi H.T., and Simons D.J., Weaker feedforward inhibition accounts for less pronounced thalamocortical response transformation in mouse vs. rat barrels. J Neurophysiol, 2013. 110(10): p. 2378–92. pmid:23966677
  9. 9. Yu C., et al., Coding of object location in the vibrissal thalamocortical system. Cereb Cortex, 2015. 25(3): p. 563–77. pmid:24062318
  10. 10. Yarbus A.L., Eye Movements and Vision. 1967, New York: Plenum Press.
  11. 11. Gibson J.J., Observations on active touch. Psychol Rev, 1962. 69: p. 477–91. pmid:13947730
  12. 12. Vincent S.B., The function of vibrissae in the behavior of the white rat. Behavior Monographs, 1912. 1(5): p. 1–82.
  13. 13. Diamond M.E., et al., 'Where' and 'what' in the whisker sensorimotor system. Nat Rev Neurosci, 2008. 9(8): p. 601–12. pmid:18641667
  14. 14. Sofroniew N.J. and Svoboda K., Whisking. Curr Biol, 2015. 25(4): p. R137–40. pmid:25689904
  15. 15. Hulliger M., et al., The responses of afferent fibres from the glabrous skin of the hand during voluntary finger movements in man. J Physiol, 1979. 291: p. 233–49. pmid:480210
  16. 16. Fee M.S., Mitra P.P., and Kleinfeld D., Central versus peripheral determinants of patterned spike activity in rat vibrissa cortex during whisking. J Neurophysiol, 1997. 78(2): p. 1144–9. pmid:9307141
  17. 17. Khatri V., et al., Whisking in air: encoding of kinematics by VPM neurons in awake rats. Somatosens Mot Res, 2010. 27(3): p. 111–20. pmid:20722492
  18. 18. Moore J.D., et al., Vibrissa Self-Motion and Touch Are Reliably Encoded along the Same Somatosensory Pathway from Brainstem through Thalamus. PLoS Biol, 2015. 13(9): p. e1002253. pmid:26393890
  19. 19. Yu C., et al., Parallel thalamic pathways for whisking and touch signals in the rat. PLoS biology, 2006. 4(5): p. e124. pmid:16605304
  20. 20. Poulet J.F., et al., Thalamic control of cortical states. Nature neuroscience, 2012. 15(3): p. 370–2. pmid:22267163
  21. 21. Urbain N., et al., Whisking-Related Changes in Neuronal Firing and Membrane Potential Dynamics in the Somatosensory Thalamus of Awake Mice. Cell Rep, 2015. 13(4): p. 647–56. pmid:26489463
  22. 22. Papakostopoulos D., Cooper R., and Crow H.J., Inhibition of cortical evoked potentials and sensation by self-initiated movement in man. Nature, 1975. 258(5533): p. 321–4. pmid:1196355
  23. 23. Giblin D.R., Somatosensory Evoked Potentials in Healthy Subjects and in Patients with Lesions of the Nervous System. Ann N Y Acad Sci, 1964. 112: p. 93–142. pmid:14188117
  24. 24. Chapman C.E., Active versus passive touch: factors influencing the transmission of somatosensory signals to primary somatosensory cortex. Can J Physiol Pharmacol, 1994. 72(5): p. 558–70. pmid:7954086
  25. 25. Fujita M., Adaptive filter model of the cerebellum. Biol Cybern, 1982. 45(3): p. 195–206. pmid:7171642
  26. 26. Kleinfeld D., et al., Adaptive filtering of vibrissa input in motor cortex of rat. Neuron, 2002. 34(6): p. 1021–34. pmid:12086648
  27. 27. Ma Y., Hu H., and Agmon A., Short-term plasticity of unitary inhibitory-to-inhibitory synapses depends on the presynaptic interneuron subtype. Journal Nuroscience, 2012. 32(3): p. 983–8. pmid:22262896
  28. 28. Lefort S., et al., The excitatory neuronal network of the C2 barrel column in mouse primary somatosensory cortex. Neuron, 2009. 61(2): p. 301–16. pmid:19186171
  29. 29. Hooks B.M., et al., Laminar analysis of excitatory local circuits in vibrissal motor and sensory cortical areas. PLoS Biol, 2011. 9(1): p. e1000572. pmid:21245906
  30. 30. Bureau I., von Saint Paul F., and Svoboda K., Interdigitated Paralemniscal and Lemniscal Pathways in the Mouse Barrel Cortex. PLoS Biol, 2006. 4(12): p. e382. pmid:17121453
  31. 31. Kinnischtzke A.K., Simons D.J., and Fanselow E.E., Motor Cortex Broadly Engages Excitatory and Inhibitory Neurons in Somatosensory Barrel Cortex. Cerebral Cortex, 2014. 24(8): p. 2237–2248. pmid:23547136
  32. 32. Porter J.T., Johnson C.K., and Agmon A., Diverse types of interneurons generate thalamus-evoked feedforward inhibition in the mouse barrel cortex. J Neurosci, 2001. 21(8): p. 2699–710. pmid:11306623
  33. 33. Bruno R.M. and Simons D.J., Feedforward mechanisms of excitatory and inhibitory cortical receptive fields. The Journal of neuroscience, 2002. 22(24): p. 10966–75. pmid:12486192
  34. 34. Cruikshank S.J., Lewis T.J., and Connors B.W., Synaptic basis for intense thalamocortical activation of feedforward inhibitory cells in neocortex. Nat Neurosci, 2007. 10(4): p. 462–8. pmid:17334362
  35. 35. Gabernet L., et al., Somatosensory integration controlled by dynamic thalamocortical feed-forward inhibition. Neuron, 2005. 48(2): p. 315–27. pmid:16242411
  36. 36. Hires S.A., et al., Low-noise encoding of active touch by layer 4 in the somatosensory cortex. eLife, 2015. 4:e06619. pmid:26245232
  37. 37. Yu J., et al., Layer 4 fast-spiking interneurons filter thalamocortical signals during active somatosensation. Nat Neurosci, 2016. 19(12): p. 1647–1657. pmid:27749825
  38. 38. van Vreeswijk C. and Sompolinsky H., Chaos in neuronal networks with balanced excitatory and inhibitory activity. Science, 1996. 274(5293): p. 1724–6. pmid:8939866
  39. 39. van Vreeswijk C. and Sompolinsky H., Chaotic balanced state in a model of cortical circuits. Neural computation, 1998. 10(6): p. 1321–71. pmid:9698348
  40. 40. Pehlevan C. and Sompolinsky H., Selectivity and sparseness in randomly connected balanced networks. PLoS One, 2014. 9(2): p. e89992. pmid:24587172
  41. 41. Tsodyks M.V., et al., Paradoxical effects of external modulation of inhibitory interneurons. J Neurosci, 1997. 17(11): p. 4382–8. pmid:9151754
  42. 42. Ozeki H., et al., Inhibitory stabilization of the cortical network underlies visual surround suppression. Neuron, 2009. 62(4): p. 578–92. pmid:19477158
  43. 43. Oh S.W., et al., A mesoscale connectome of the mouse brain. Nature, 2014. 508(7495): p. 207–14. pmid:24695228
  44. 44. Petreanu L., et al., The subcellular organization of neocortical excitatory connections. Nature, 2009. 457(7233): p. 1142–1145. pmid:19151697
  45. 45. Petreanu L., et al., Channelrhodopsin-2-assisted circuit mapping of long-range callosal projections. Nat Neurosci, 2007. 10(5): p. 663–8. pmid:17435752
  46. 46. Jiang X., et al., Principles of connectivity among morphologically defined cell types in adult neocortex. Science, 2015. 350(6264): p. aac9462. pmid:26612957
  47. 47. Knutsen P.M., Pietr M., and Ahissar E., Haptic object localization in the vibrissal system: behavior and performance. J Neurosci, 2006. 26(33): p. 8451–64. pmid:16914670
  48. 48. O'Connor D.H., et al., Vibrissa-based object localization in head-fixed mice. J Neurosci, 2010. 30(5): p. 1947–67. pmid:20130203
  49. 49. Clack N.G., et al., Automated tracking of whiskers in videos of head fixed rodents. PLoS computational biology, 2012. 8(7): p. e1002591. pmid:22792058
  50. 50. Pammer L., et al., The Mechanical Variables Underlying Object Localization along the Axis of the Whisker. J Neurosci, 2013. 33(16): p. 6726–41. pmid:23595731
  51. 51. O'Connor D.H., et al., Neural coding during active somatosensation revealed using illusory touch. Nat Neurosci, 2013. 16(7): p. 958–65. pmid:23727820
  52. 52. Fanselow E.E., et al., Thalamic bursting in rats during different awake behavioral states. Proc Natl Acad Sci U S A, 2001. 98(26): p. 15330–5. pmid:11752471
  53. 53. Bruno R.M. and Sakmann B., Cortex is driven by weak but synchronously active thalamocortical synapses. Science, 2006. 312(5780): p. 1622–7. pmid:16778049
  54. 54. Han X. and Boyden E.S., Multiple-color optical activation, silencing, and desynchronization of neural activity, with single-spike temporal resolution. PLoS ONE, 2007. 2: p. e299. pmid:17375185
  55. 55. Zhang F., et al., Multimodal fast optical interrogation of neural circuitry. Nature, 2007. 446(7136): p. 633–9. pmid:17410168
  56. 56. Beierlein M., Gibson J.R., and Connors B.W., Two dynamically distinct inhibitory networks in layer 4 of the neocortex. J Neurophysiol, 2003. 90(5): p. 2987–3000. pmid:12815025
  57. 57. Feldmeyer D., Roth A., and Sakmann B., Monosynaptic connections between pairs of spiny stellate cells in layer 4 and pyramidal cells in layer 5A indicate that lemniscal and paralemniscal afferent pathways converge in the infragranular somatosensory cortex. J Neurosci, 2005. 25(13): p. 3423–31. pmid:15800197
  58. 58. Hansel D. and van Vreeswijk C., The mechanism of orientation selectivity in primary visual cortex without a functional map. J Neurosci, 2012. 32(12): p. 4049–64. pmid:22442071
  59. 59. Wang X.J. and Buzsaki G., Gamma oscillation by synaptic inhibition in a hippocampal interneuronal network model. J Neurosci, 1996. 16(20): p. 6402–13. pmid:8815919
  60. 60. Hull C., Isaacson J.S., and Scanziani M., Postsynaptic mechanisms govern the differential excitation of cortical neurons by thalamic inputs. J Neurosci, 2009. 29(28): p. 9127–36. pmid:19605650
  61. 61. Pinto D.J., et al., Cortical damping: analysis of thalamocortical response transformations in rodent barrel cortex. Cereb Cortex, 2003. 13(1): p. 33–44. pmid:12466213
  62. 62. Wilent W.B. and Contreras D., Dynamics of excitation and inhibition underlying stimulus selectivity in rat somatosensory cortex. Nat Neurosci, 2005. 8(10): p. 1364–70. pmid:16158064
  63. 63. Kremkow J., et al., Functional consequences of correlated excitatory and inhibitory conductances in cortical networks. J Comput Neurosci, 2010. 28(3): p. 579–94. pmid:20490645
  64. 64. Pinto D.J., Brumberg J.C., and Simons D.J., Circuit dynamics and coding strategies in rodent somatosensory cortex. J Neurophysiol, 2000. 83(3): p. 1158–66. pmid:10712446
  65. 65. Huber D., et al., Sparse optical microstimulation in barrel cortex drives learned behaviour in freely moving mice. Nature, 2008. 451(7174): p. 61–4. pmid:18094685
  66. 66. Rubin D.B., Van Hooser S.D., and Miller K.D., The stabilized supralinear network: a unifying circuit motif underlying multi-input integration in sensory cortex. Neuron, 2015. 85(2): p. 402–17. pmid:25611511
  67. 67. O'Connor D.H., et al., Neural activity in barrel cortex underlying vibrissa-based object localization in mice. Neuron, 2010. 67(6): p. 1048–61. pmid:20869600
  68. 68. Agmon A. and Connors B.W., Thalamocortical responses of mouse somatosensory (barrel) cortex in vitro. Neuroscience, 1991. 41(2–3): p. 365–79. pmid:1870696
  69. 69. Wehr M. and Zador A.M., Balanced inhibition underlies tuning and sharpens spike timing in auditory cortex. Nature, 2003. 426(6965): p. 442–6. pmid:14647382
  70. 70. DeWeese M.R., Wehr M., and Zador A.M., Binary spiking in auditory cortex. J Neurosci, 2003. 23(21): p. 7940–9. pmid:12944525
  71. 71. Hayut I., et al., LTS and FS inhibitory interneurons, short-term synaptic plasticity, and cortical circuit dynamics. PLoS Comput Biol, 2011. 7(10): p. e1002248. pmid:22046121
  72. 72. Markram H., et al., Reconstruction and Simulation of Neocortical Microcircuitry. Cell, 2015. 163(2): p. 456–92. pmid:26451489
  73. 73. Traub R.D., Miles R., and Wong R.K., Model of the origin of rhythmic population oscillations in the hippocampal slice. Science, 1989. 243(4896): p. 1319–25. pmid:2646715
  74. 74. Eliasmith C. and Trujillo O., The use and abuse of large-scale brain models. Curr Opin Neurobiol, 2014. 25: p. 1–6. pmid:24709593
  75. 75. Marder E. and Goaillard J.M., Variability, compensation and homeostasis in neuron and network function. Nat Rev Neurosci, 2006. 7(7): p. 563–74. pmid:16791145
  76. 76. Shriki O., Hansel D., and Sompolinsky H., Rate models for conductance-based cortical neuronal networks. Neural Comput, 2003. 15(8): p. 1809–41. pmid:14511514
  77. 77. Reinhold K., Lien A.D., and Scanziani M., Distinct recurrent versus afferent dynamics in cortical visual processing. Nat Neurosci, 2015. 18(12): p. 1789–97. pmid:26502263
  78. 78. Bruno R.M., Synchrony in sensation. Current opinion in neurobiology, 2011. 21(5): p. 701–8. pmid:21723114
  79. 79. Reyes A.D., Synchrony-dependent propagation of firing rate in iteratively constructed networks in vitro. Nat Neurosci, 2003. 6(6): p. 593–9. pmid:12730700
  80. 80. Lerchner A., et al., Mean field theory for a balanced hypercolumn model of orientation selectivity in primary visual cortex. Network, 2006. 17(2): p. 131–50. pmid:16818394
  81. 81. Hansel D. and Mato G., Short-term plasticity explains irregular persistent activity in working memory tasks. J Neurosci, 2013. 33(1): p. 133–49. pmid:23283328
  82. 82. Ahmadian Y., Rubin D.B., and Miller K.D., Analysis of the stabilized supralinear network. Neural Comput, 2013. 25(8): p. 1994–2037. pmid:23663149
  83. 83. Persi E., et al., Power-law input-output transfer functions explain the contrast-response and tuning properties of neurons in visual cortex. PLoS Comput Biol, 2011. 7(2): p. e1001078. pmid:21390280
  84. 84. Curtis J.C. and Kleinfeld D., Phase-to-rate transformations encode touch in cortical neurons of a scanning sensorimotor system. Nat Neurosci, 2009. 12(4): p. 492–501. pmid:19270688
  85. 85. Birdwell J.A., et al., Biomechanical models for radial distance determination by the rat vibrissal system. J Neurophysiol, 2007. 98(4): p. 2439–55. pmid:17553946
  86. 86. Song S., et al., Highly nonrandom features of synaptic connectivity in local cortical circuits. PLoS Biol, 2005. 3(3):e68 p. 1–13. pmid:15737062
  87. 87. Ma Y., et al., Distinct subtypes of somatostatin-containing neocortical interneurons revealed in transgenic mice. J Neurosci, 2006. 26(19): p. 5069–82. pmid:16687498
  88. 88. Ferster D., Chung S., and Wheat H., Orientation selectivity of thalamic input to simple cells of cat visual cortex. Nature, 1996. 380: p. 249–252. pmid:8637573
  89. 89. Golomb D., et al., On temporal codes and the spatiotemporal response of neurons in the lateral geniculate nucleus. J Neurophysiol, 1994. 72(6): p. 2990–3003. pmid:7897504
  90. 90. Golomb D. and Hansel D., The number of synaptic inputs and the synchrony of large, sparse neuronal networks. Neural Comput, 2000. 12(5): p. 1095–139. pmid:10905810
  91. 91. Varshney L.R., Sjostrom P.J., and Chklovskii D.B., Optimal information storage in noisy synapses under resource constraints. Neuron, 2006. 52(3): p. 409–23. pmid:17088208
  92. 92. Cossell L., et al., Functional organization of excitatory synaptic strength in primary visual cortex. Nature, 2015. 518(7539): p. 399–403. pmid:25652823
  93. 93. Alfonsa H., et al., The contribution of raised intraneuronal chloride to epileptic network activity. J Neurosci, 2015. 35(20): p. 7715–26. pmid:25995461
  94. 94. Raimondo J.V., et al., Optogenetic silencing strategies differ in their effects on inhibitory synaptic transmission. Nat Neurosci, 2012. 15(8): p. 1102–4. pmid:22729174
  95. 95. Zillmer R., Brunel N., and Hansel D., Very long transients, irregular firing, and chaotic dynamics in networks of randomly connected inhibitory integrate-and-fire neurons. Phys Rev E Stat Nonlin Soft Matter Phys, 2009. 79(3 Pt 1): p. 031909. pmid:19391973