Synaptic and Network Mechanisms of Sparse and Reliable Visual Cortical Activity during Nonclassical Receptive Field Stimulation

SUMMARY

During natural vision, the entire visual field is stimulated by images rich in spatiotemporal structure. Although many visual system studies restrict stimuli to the classical receptive field (CRF), it is known that costimulation of the CRF and the surrounding nonclassical receptive field (nCRF) increases neuronal response sparseness. The cellular and network mechanisms underlying increased response sparseness remain largely unexplored. Here we show that combined CRF + nCRF stimulation increases the sparseness, reliability, and precision of spiking and membrane potential responses in classical regular spiking (RS_C) pyramidal neurons of cat primary visual cortex. Conversely, fast-spiking interneurons exhibit increased activity and decreased selectivity during CRF + nCRF stimulation. The increased sparseness and reliability of RS_C neuron spiking is associated with increased inhibitory barrages and narrower visually evoked synaptic potentials. Our experimental observations were replicated with a simple computational model, suggesting that network interactions among neuronal subtypes ultimately sharpen recurrent excitation, producing specific and reliable visual responses.

INTRODUCTION

Is the cortical code for sensory information redundant, or is it sparse and efficient? One influential theory proposes that sensory systems use highly selective stimulus representations that optimize metabolic efficiency by minimizing the number of action potentials (Barlow, 1972). Numerous studies have shown that appropriate stimuli (e.g., high-contrast bright or dark bars) placed in a restricted portion of the visual field increase the firing rates of neurons in primary visual cortex; this area of maximal sensitivity is termed the minimum response field or classical receptive field (CRF; DeAngelis et al., 1993; Hubel and Wiesel, 1962; Movshon et al., 1978). However, stimulation of regions adjacent to the CRF (collectively termed the nonclassical receptive field, nCRF), where the same stimuli fail to elicit spikes, can modulate responses to CRF stimuli in complex and often nonlinear ways. Typically, costimulation of the CRF + nCRF suppresses spiking responses compared with CRF stimulation alone, although there are also examples of nCRF-mediated contextual enhancement (Angelucci and Bressloff, 2006; Cava-naugh et al., 2002a; Fitzpatrick, 2000; Gilbert et al., 1996; Jones et al., 2001; Kapadia et al., 1995; Webb et al., 2005). Integrating the modulatory properties of the nCRF with established properties of the CRF into a single model framework is an essential step toward a complete understanding of cortical information coding and neuronal responsiveness (see Carandini et al., 2005).

Recent studies in awake, behaving primates, motivated in part by the efficient coding hypothesis, have suggested an alternative framework for considering nCRF modulation of CRF activity. During naturalistic visual stimulation of both the CRF + nCRF, spiking responses in primary visual cortex are neither uniformly suppressed nor facilitated, but instead display nonlinear modulations resulting in a net increase in response sparseness (Vinje and Gallant, 2000). In these studies and others, response sparseness serves as a proxy for neural selectivity: a neuron with increased sparseness responds to a more restricted set of stimuli, and is thus more selective across the entire stimulus set (Lehky et al., 2005; Olshausen and Field, 2004; Rolls and Tovee, 1995; Tolhurst et al., 2009; Yao et al., 2007; Yen et al., 2007). Compared with stimulation of the CRF alone, sparser spike responses elicited by wide-field stimulation contain specific epochs of both suppression and facilitation that, as a whole, transmit more information about the stimulus (Vinje and Gallant, 2002). However, it remains unclear how sparse single-neuron responses can be reliably transmitted to downstream neurons in the face of typical trial-to-trial response variability (Shadlen and Newsome, 1998; Stein et al., 2005). It is therefore critical to determine how the cortical network overcomes this inherent response variability in order to generate and transmit sparse neuronal responses during wide-field stimulation.

One important observation that may help explain the origins of sparse sensory coding is that the amplitude and timing of action potentials depends critically upon the connectivity and activity levels of presynaptic excitatory and inhibitory neuronal subtypes (Azouz et al., 1997; Bruno and Sakmann, 2006; Contreras and Palmer, 2003; Silberberg and Markram, 2007; Swadlow, 2003; Yoshimura and Callaway, 2005). The net pattern of activity across the distributed network of presynaptic neuronal subpopulations is collectively visible in a single neuron’s membrane potential fluctuations, where the amplitude and precise timing of excitatory and especially inhibitory potentials is a critical factor in determining exactly when a given pyramidal neuron spikes (Gabernet et al., 2005; Haider and McCormick, 2009; Hasen-staub et al., 2005; Higley and Contreras, 2006; Pouille and Scan-ziani, 2001; Wehr and Zador, 2003).

Here, we report that wide-field naturalistic stimulation of classical regular spiking (RS_C) pyramidal neurons in cat primary visual cortex both increased the amplitude of inhibitory postsyn-aptic potentials (IPSPs) and increased the trial-to-trial reliability of excitatory postsynaptic potentials (EPSPs). These synaptic events were mirrored by spiking activity recorded in specific excitatory and inhibitory neuronal subpopulations. Injection of these recorded PSP sequences into a simple computational model revealed that while both changes in IPSPs and EPSPs contributed to the sparseness and reliability of spiking responses during CRF + nCRF stimulation, IPSPs were the primary cause of increased neuronal sparseness, while the EPSPs greatly enhanced trial-to-trial spike reliability and predominantly shaped the overall spiking response. These combined excitatory and inhibitory network interactions may ultimately explain the highly selective and more reliable neuronal activity observed in visual cortex in response to naturalistic sensory inputs.

RESULTS

Stimulus Presentation and Experimental Approach

We performed extracellular and intracellular recordings in cat primary visual cortex during presentation of naturalistic movie sequences masked by circular apertures of two sizes. One restricted the movie to the CRF (as defined quantitatively, see Experimental Procedures and Supplemental Experimental Procedures available online), while the other was larger and resulted in costimulation of both the CRF and the surrounding nCRF. Note that in both configurations visual stimulation within the CRF was identical—only the size of the aperture was changed (from a radius of 1X CRF to 3X CRF).

We first present results obtained from electrophysiologically identified classical regular spiking (RS_C) neurons, the most abundant and most frequently recorded cortical neuronal subtype (Nowak et al., 2003).

CRF + nCRF Stimulation of RS_C Neurons Increases Response Sparseness and Membrane Hyperpolarization

Isolated CRF stimulation in a typical RS_C neuron elicited robust action potentials across repeated presentations of the stimulus (Figures 1A and 1C). Both the temporal structure and amplitude of this neuron’s response changed dramatically in response to the larger CRF + nCRF stimulus (Figures 1B and 1D), including a reduction in amplitude and change in stimulus frame eliciting a maximal response (compare black and red arrowheads Figures 1C and 1D). In addition, this neuron was markedly hyperpolarized during CRF + nCRF stimulation compared with CRF alone stimulation (dashed lines in Figures 1A and 1B), while depolarizing synaptic activity was sharper and more reliable across repeated presentations during CRF + nCRF stimulation (Figure 1B), compared with CRF-alone stimulation (Figure 1A). As a result, the neuron responded to fewer movie frames and spiking responses were more transient and temporally precise. This neuron exhibited a net decrease in responsiveness throughout the CRF + nCRF movie, although some peaks in the peristimulus time spike histogram (PSTH) were only mildly reduced (Figure 1C, closed arrows) while others were nearly abolished (Figure 1C, open arrows). In other RS_C neurons, the effects of CRF + nCRF stimulation were more complex: some frames elicited greater spiking activity in the CRF + nCRF con-figuration, while other frames elicited reductions (Figure S1). This finding, consistent with previous studies in awake, behaving primates (Vinje and Gallant, 2000), suggests that dynamic, spatiotemporally rich CRF + nCRF costimulation engages more complex mechanisms than uniform response suppression.

Figure 1. Naturalistic Wide-Field Visual Stimulation Increases Selectivity.

(A) Intracellular responses of an RS_C neuron to repeated presentations (five) of a natural scene movie restricted to the classical receptive field (CRF). Average membrane potential (V_m) = −57.8 mV. Inset shows extent of the movie overlying the CRF; mask was opaque during recordings. The selectivity or sparseness index (S) was 0.29 ± 0.01 (mean and standard error of the mean [± SEM] throughout).

(B) Responses to five repeats of the same movie with a larger aperture that stimulated portions of the nonclas-sical receptive field (nCRF) in addition to the CRF. Average V_m = −65.7 mV. Sparseness increased to 0.72 ± 0.01. See also Movie S1.

(C and D) Histograms of spiking responses to CRF stimulation (black) and (D) combined CRF + nCRF stimulation (red). Peak CRF response to best frame (45.9 Hz; black arrowhead) occurs 1.4 s after movie onset. Peak CRF + nCRF response (17.2 Hz; red arrowhead) occurs 0.6 s after movie onset. Histograms appear twice (C and D) and are overlaid to facilitate comparison. Note that CRF + nCRF costimulation results in the suppression of some peaks present in the CRF response (open arrows), while others are less affected (closed arrows). See also Figure S1.

(E) Spiking responses became significantly more sparse (see text) in all 13 neurons (inset), corresponding to a 23% net increase in sparseness with combined CRF + nCRF stimulation (S_{CRF + nCRF} = 0.69 ± 0.02) compared with CRF alone stimulation (S_CRF = 0.56 ± 0.02; p < 0.01) across the population of RS_C neurons. See also Tables S1 and S2.

(F) Neurons were significantly hyperpolarized (−1.6 mV on average; 10/13 individually, inset) during CRF + nCRF stimulation (V_{m CRF + nCRF} = −65.3 ± 0.4 mV; V_{m CRF} =−63.7 ± 0.6 mV; p < 0.01) in comparison to CRF only stimulation.

To quantify these complex changes in responsiveness, we compared the lifetime sparseness of spike responses in the two stimulation conditions (see Experimental Procedures; see also Vinje and Gallant, 2000; Willmore and Tolhurst, 2001). Sparseness is a nonparametric measure of neuronal selectivity that is a property of both the neuron and the stimulus set. For a fixed stimulus set, sparseness can be computed without knowledge of a neuron’s classical tuning properties (i.e., orientation, spatial frequency, and phase tuning) and thus provides a robust metric for quantifying changes in selectivity without requiring assumptions about, or complete measurements of, the underlying spatiotemporal tuning profile of either the CRF or the nCRF. The sparseness metric (S) we use here is bounded between 0 and 1 (Vinje and Gallant, 2000); S approaches 0 for nonselective neurons that respond equally to all movie frames (dense coding) and approaches 1 for highly selective neurons, where, in the limiting case, spikes are elicited by a single movie frame (sparse coding). CRF + nCRF stimulation resulted in a significantly sparser response compared with CRF alone stimulation in all RS_C neurons (Figure 1E; n = 13), with a population average 23% increase in lifetime sparseness (Figure 1E; see figure legends for all p values throughout). Importantly, we observed the same pattern of increased sparseness in our extra-cellular multiunit (MU) recording data (18% average increase in sparseness; n = 16 MU recordings; Table S1), and we confirmed that this effect was not simply due to CRF/nCRF boundary stimulation (Table S1).

Interestingly, we observed a significant hyperpolarization in 10/13 RS_C neurons (mean −1.6 mV) during CRF + nCRF stimulation (Figure 1F; n = 13). Hyperpolarization appears to contribute in part to increased selectivity, because hyperpolarizing neurons with direct negative current injection during CRF movie presentation also increased response selectivity. However, on average, the rate of increase in sparseness with DC injection-induced hyperpolarization during CRF-alone stimulation was much less than that occurring during natural CRF + nCRF hyperpolarization (0.07 ΔS per mV DC hyperpolarization, versus 0.22 ΔS per mV with natural CRF + nCRF hyperpolarization; n = 10 RS_C neurons; p < 0.01; data not shown; see Supplemental Information), suggesting that additional mechanisms beyond simple hyperpolarization must be involved in increasing neuronal selectivity.

Isolated nCRF Stimulation of RS_C Neurons Does Not Increase Selectivity or Elicit Membrane Potential Hyperpolarization

We found that nCRF stimulation alone (annulus) cannot account for the observed effects, because it did not significantly increase average firing rates above spontaneous activity, in either intra-cellularly recorded RS_C neurons or in MU responses (Table S2), nor did it result in a net change in the average membrane potential compared with a blank screen (Table S2). However, compared with spontaneous activity, there was a significant increase in membrane potential standard deviation in RS_C cells during nCRF stimulation, indicative of increased synaptic activity, consistent with previous reports (Monier et al., 2003).

Taken together, these observations indicate that nCRF costimulation counterbalances much of the strong average depolarization associated with CRF stimulation, either through a reduction in visually evoked EPSP, or an increase in IPSP, barrages.

CRF + nCRF Stimulation Increases the Amplitude of IPSP Barrages

To examine the specific contributions of excitation and inhibition to visual selectivity in RS_C neurons, we pharmacologically blocked several intrinsic neuronal conductances by including QX-314 (blocks Na⁺ currents and the h-current) and Cs⁺ (blocks K⁺ currents) in the intracellular recording electrode. We then compared EPSPs and IPSPs evoked by CRF and CRF + nCRF stimulation in neurons tonically hyperpolarized to around −75 mV (near the reversal potential for GABAergic Cl⁻ inhibition) or depolarized to around 0 mV (near the reversal potential for glutamatergic excitation) with intracellular current injection (see Experimental Procedures). Stimulation of the CRF + nCRF resulted in a large increase in average IPSP amplitude relative to CRF alone stimulation (Figure 2A, blue trace; increase is downward for IPSPs), with little or no effect on average EPSP amplitude (Figure 2B). Across the population of RS_C neurons (n = 9), we found a significant increase in the overall amplitude of evoked IPSPs (Figure 2C, blue, ΔIPSP = −1.99 ± 0.4 mV; 43.7 ± 12.0% increase) and no significant difference in the overall amplitude of EPSPs evoked by CRF + nCRF stimulation compared with CRF alone stimulation (Figure 2C, red, ΔEPSP = 0.33 ± 0.2 mV; 4.9 ± 4.3% increase).

Figure 2. Wide-Field Visual Stimulation Selectively Increases the Amplitude of Inhibitory Postsynaptic Potentials.

(A) Average inhibitory postsynaptic potentials (IPSPs) recorded during 12 presentations of a naturalistic movie to the CRF (black traces) and to the CRF + nCRF (blue). QX-314 and Cs⁺ in micropipette. Upper and lower thin traces indicate ± SEM. Dashed vertical line indicates movie onset, downward deflections indicate IPSPs. Some IPSP barrages increase greatly (closed arrows), while others change little (open arrows). CRF + nCRF stimulation significantly increases average IPSP amplitude (compared with average response during CRF stimulation) in this cell by −2.73 ± 0.51 mV (p < 0.01; recorded at 0 mV).

(B) In contrast to IPSPs, EPSPs (recorded at −75 mV) did not significantly differ in amplitude between the two conditions (−0.41 ± 0.61 mV; p > 0.1). The neuron shown here is the same neuron as shown in Figures 1A and 1B after spike inactivation.

(C) Population differences in EPSPs (red) and IPSPs (blue) evoked with CRF + nCRF stimulation, compared with CRF alone stimulation. All nine neurons were determined to be RS_C before spike inactivation. CRF + nCRF evoked IPSP barrages were significantly larger on average (blue, −1.99 ± 0.4 mV; 43.7 ± 12.0% increase; p < 0.01) while EPSP barrages were not (0.33 ± 0.2 mV; 4.9 ± 4.3% increase; p > 0.1). Values are mean ± SEM.

CRF + nCRF Stimulation of Fast-Spiking Interneurons Decreases Selectivity and Increases Response Amplitude

The finding of increased visually evoked IPSPs barrages in RS_C neurons during CRF + nCRF stimulation suggests that this change is partially driven by increased activity in one or more subtypes of cortical inhibitory neurons. The only subtype of inhibitory neuron unambiguously identifiable under our recording conditions is the fast-spiking (FS) interneuron (Nowak et al., 2003). Recordings from electrophysiologically identified FS cells (Figure 3C inset) revealed increased firing rates and decreased response sparseness during CRF + nCRF stimulation (Figures 3A–3C and 4E). As with changes in RS_C spiking activity, CRF + nCRF stimulation elicited nonlinear changes in FS inter-neuron activity (Figures 3B and 3C open arrows), as echoed by nonlinear modulation of IPSPs recorded in RS_C neurons (see Supplemental Information).

Figure 3. Fast-Spiking Interneurons and Thin-Spike Regular-Spiking Neurons Become More Active and Less Sparse during CRF + nCRF Stimulation.

(A) Intracellular responses of an electrophysiologically identified FS interneuron (inset, shows sustained firing rate >300 Hz in response to current pulse) during ten trials of CRF stimulation (black).

(B) CRF + nCRF stimulation (red) elicits larger responses, compared with the CRF configuration (closed arrows).

(C) PSTHs from 15 repeated trials of CRF (black) and CRF + nCRF presentations (red) reveal elevated PSTH peaks (closed arrows), and the appearance of new peaks (open arrow) during wide-field stimulation. FS interneuron population (n = 5 intracellular, n = 4 extracellular) significantly decreased response sparseness (12%) with CRF + nCRF stimulation (S_CRF = 0.48 ± 0.007; S_{CRF + nCRF} = 0.43 ± 0.007; p < 0.01). Values are mean ± SEM.

(D) Intracellular response of an RS_TS neuron (inset, adapting firing pattern to current pulse, rate −100 Hz, spike width at half height 0.25 ms) during five trials of CRF stimulation (black).

(E) Response of same neuron to five trials of CRF + nCRF stimulation (red). Note increased action potential response (closed arrows) and addition of new responses (open arrow).

(F) PSTH across 15 trials of CRF stimulation (black) and CRF + nCRF stimulation (red) reveals elevated PSTH peaks (closed arrow), along with addition of peaks (open arrows) during wide-field stimulation. Inset, RS_TS neuron population (n = 12 intracellular, 3 juxtacellular) significantly decreased sparseness (7% average decrease; S_CRF = 0.66 ± 0.006; S_{CRF + nCRF} = 0.62 ± 0.005; p < 0.01) during CRF + nCRF stimulation. See Figure S2 for RS_TS neurons, and Table S3 for biophysical and functional response properties of cell classes. Values are mean ± SEM.

Figure 4. Correlated Activity in Subthreshold and Spiking Responses in Distinct Excitatory Networks Drives Increased Reliability of Visual Responses during Wide-Field CRF + nCRF Stimulation.

(A) Response of an RS_C neuron to five natural movie presentations to the CRF (current pulse response, inset). Note the trial-to-trial variability of membrane potential (V_m) response.

(B) Same neuron, responses to five trials of CRF + nCRF movie. Across all 20 trials, there was a 21% increase in the reliability of V_m across trials (inset; R_Vm = 0.68) and a 163% increase in reliability of spike responses (R_Spikes = 0.29) compared with CRF stimulation (inset in 4A; R_Vm = 0.56; R_Spikes = 0.11; p < 0.01 for both comparisons). Sparseness across all trials also significantly increased (p < 0.01).

(C) Trial-to-trial membrane potential response reliability of RS_C neuron population (n = 13) significantly increases with CRF + nCRF stimulation (V_m R_CRF = 0.26 ± 0.01; V_m R_{CRF + nCRF} = 0.31 ± 0.01; p < 0.01) in parallel with increased reliability of spike responses in these same neurons (spikes R_CRF = 0.12 ± 0.01; spikes R_{CRF + nCRF} = 0.18 ± 0.01; p < 0.01). Values are mean ± SEM. See also Figure S4 for similar results in MU recordings.

(D) Isolated EPSPs significantly increase reliability (by 70%) with CRF + nCRF stimulation (EPSP R_CRF = 0.16 ± 0.01; EPSP R_{CRF + nCRF} = 0.27 ± 0.02; p < 0.01), while IPSP reliability does not significantly change (IPSP R_CRF = 0.14 ± 0.01; IPSP R_{CRF + nCRF} = 0.13 ± 0.01; p > 0.1).

(E) Normalized firing rates of both FS and RS_TS neurons increase significantly with CRF + nCRF stimulation (22.8 ± 6.3% and 26.8 ± 12.5%, respectively; p < 0.01, sign test), while normalized firing rates of RS_C neurons decrease significantly with CRF + nCRF stimulation (−21.2 ± 13.4%; p < 0.01, sign test). Firing rates normalized to CRF alone average firing rates for each neuron (FS: 7.6 ± 1.8 Hz; RS_TS: 2.8 ± 0.7 Hz; RS_C: 1.8 ± 1.2 Hz).

(F) RS_TS neurons significantly decrease their spike-train reliability (black and red, left) with CRF + nCRF stimulation, (R_CRF = 0.37 ± 0.02; R_{CRF + nCRF} = 0.30 ± 0.02; p < 0.01) while FS neurons maintain high spike-train reliability (black and blue, right) with CRF + nCRF stimulation, (R_CRF = 0.29 ± 0.02; R_{CRF + nCRF} = 0.28 ± 0.02; p > 0.1).

CRF + nCRF Stimulation of Thin-Spike Regular-Spiking Neurons, like FS Interneurons, Results in Decreased Response Sparseness and Increased Response Amplitude

While searching for FS interneurons, we preferentially recorded from neurons with thin action potentials, this being one of several—but not the only—defining characteristic of FS interneurons (Nowak et al., 2003). As a result, we also recorded from a substantial fraction of neurons with unusually thin action potentials that nonetheless exhibited spike frequency adaptation to current pulse injection (Figure 4D, inset) with sustained firing rates <200 Hz during the pulse. These neurons have previously been shown to belong to a subclass of regular spiking neurons termed thin-spike regular-spiking neurons (RS_TS; Figure S2; Nowak et al.,2003). Intracellular labeling of these neurons revealed them to be spiny pyramidal neurons (Figure S2), consistent with previous reports (Nowak et al., 2003). RS_TS cells are distinct from chattering neurons (which also display thin spikes) in that they do not discharge intrinsic bursts of action potentials (Nowak et al., 2003). The majority (11/15 or 73%) of RS_TS neurons, like FS interneurons, showed significantly increased firing rates (see below) and decreased response sparseness during CRF + nCRF stimulation (Figures 3D–3F).

FS and RS_TS Neurons Are Functionally Distinct from RS_C Neurons

Not only did these three classes of cells (RS_C, FS, and RS_TS) exhibit differences in their response to wide-field visual stimulation, but they also displayed unique biophysical properties. As previously reported, both FS and RS_TS neurons exhibit significantly narrower spike widths (0.19 ± 0.01 and 0.24 ± 0.02 ms at half height, respectively) and significantly faster membrane time constants (τ) compared with RS_C neurons (Table S3; Nowak et al., 2003). In addition, the power-law relationship between visually evoked membrane potential changes and firing rate (Anderson et al., 2000; Miller and Troyer, 2002) was significantly steeper for FS neurons compared with either RS_TS or RS_C neurons (Figure S3). This predicts that FS neurons will exhibit greater nonlinear increases in firing rate for the same net depolarizing synaptic input than either RS_C or RS_TS pyramidal neurons.

Simple and complex cells were represented with similar frequency across all three cell types (Table S3). We also observed that the size of the CRF was significantly larger and the response latency significantly shorter in FS and RS_TS neurons than in RS_C neurons (Table S3). Together, these results demonstrate that a neuron’s unique biophysical properties (i.e., neuronal subtype) along with its spatial and temporal integrative properties are predictive of the response to wide-field visual stimulation.

CRF + nCRF Stimulation of RS_C Neurons Increases Trial-to-Trial Reliability of Evoked Synaptic and Action Potential Responses

Throughout our experiments, we repeatedly observed that trial-to-trial response reliability of both subthreshold membrane potential fluctuations and spike times increased during CRF + nCRF stimulation (i.e., decreased variability across repeated trials; Figures 4A and 4B; see also Figures 1A and 1B).To quantify subthreshold trial-to-trial response reliability, we performed a pair-wise cross correlation analysis of membrane potential (V_m) responses recorded on each trial versus every other trial, within each RS_C neuron, separated by experimental condition (following digital spike removal, see Supplemental Experimental Procedures). In RS_C neurons we observed an 81.4 ± 12.2% increase in mean membrane potential reliability and a 298.1 ± 45.7% increase in spike train reliability across trials during wide-field stimulation (Figure 4C). A similar pattern of increased sparseness simultaneous with increased trial-to-trial reliability was observed in our MU recordings (2-fold reliability increase; n = 16 recordings; Figure S4).

Changes in Spiking Activity Reflect Changes in the Amplitude and Reliability of EPSPs and IPSPs Evoked by Wide-Field Stimulation

We next examined whether the increase in membrane potential reliability with wide-field stimulation resulted in increased reliability of EPSPs, IPSPs, or both (see Figure 2). As expected, trial-to-trial reliability of both EPSPs and IPSPs was significantly greater than expected from temporally shuffled data (data not shown), but surprisingly, the peak IPSP reliability was unaltered between the two stimulus conditions (Figure 4D, black and blue bars). However, for these same neurons, trial-to-trial EPSP reliability significantly increased (by nearly 70%) during CRF + nCRF stimulation (Figure 4D, black and red bars).

We wondered whether the unique properties of the different neuronal subtypes discussed above (Table S3) could underlie the increase in hyperpolarization (Figure 1F) and IPSP amplitudes (Figure 2C) without an accompanying change in IPSP reliability (Figure 4D). We re-examined the activity levels in each cell type, and found the average firing rates of both FS and RS_TS neurons increased significantly during CRF + nCRF stimulation (FS: 22.8 ± 6.3%; RS_TS: 26.8 ± 12.5%; normalized by average rate during CRF stimulation; Figure 4E, red). While not significantly different from each other, these increases were significantly greater than the change in normalized firing rate for RS_C neurons, which decreased significantly (−21.2 ± 13.4%) during CRF + nCRF stimulation (Figure 4E, black).

Interestingly, like IPSP reliability in RS_C neurons, the high reliability of FS interneuron spike trains was not altered by CRF + nCRF stimulation (Figure 4F; black and blue). In addition, the spike train reliability of RS_TS neurons decreased significantly during CRF + nCRF stimulation (Figure 4F; red and blue). Note, however, that although RS_TS neurons showed a relative decrease in trial-to-trial reliability during CRF + nCRF stimulation, RS_TS neurons were consistently more reliable than RS_C neurons under similar stimulation conditions (cf. Figure 4C).

Increased Temporal Precision of RS_C Spiking Is Associated with Sharper Synaptic Responses

The results presented thus far suggest that increased EPSP reliability in RS_C neurons may be due to increased spiking reliability in other RS_C pyramidal neurons within the cortical network. However, it is also possible that increased reliability arises from increased temporal precision of spikes across the population of RS_C neurons. In many neurons, peaks in the PSTH became sharper in the CRF + nCRF condition compared with the CRF condition (e.g., Figures 1D, 4A, and 4B), suggesting an increase in temporal precision. To quantify this observation we computed the change in width of the central peak of the PSTH’s autocovariance function between the two stimulation conditions (e.g., Desbordes et al., 2008). Because the typical RS_C PSTH exhibited a few peaks interrupted by periods of relative silence (e.g., Figures 4A, 4B, and S2), this method adequately captures the temporal extent of the average PSTH event, which is the summed spike activity elicited by a few select movie frames across trials.

The average half-width of the central peak of the PSTH auto-covariance function (see Supplemental Experimental Procedures) significantly decreased during CRF + nCRF stimulation (87.5 ms; 35%) in a representative RS_C neuron (Figure 5A, inset) and by 33% in the RS_C population (Figure 5A; n = 13). The increase in PSTH precision was also associated with a reduced mean and modal interspike interval, corresponding to an increase in instantaneous firing rates for select portions of RS_C spike trains strongly driven by specific movie frames (Figure S5; see also Figure S2).

Figure 5. Temporal Precision of Spike Responses in RS_C Neurons Increases with CRF + nCRF Stimulation and Is Associated with Narrowing of the Underlying Synaptic Events.

(A) Width of the autocovariance function of a representative RS_C neuron’s PSTH is significantly (35%) narrower with combined CRF + nCRF stimulation (red) compared with CRF alone stimulation (black). Across the population of RS_C neurons (n = 13), there was a significant narrowing (by 33%) of the average event in the PSTH with combined CRF + nCRF stimulation (181.6 ± 15.6 ms, red bar) compared with CRF alone stimulation (272.4 ± 23.9 ms, black bar; p < 0.01). See also Figure S5 for interspike interval histograms. Values are mean ± SEM.

(B) Spike-triggered average of V_m in these same neurons reveals a narrower synaptic potential underlying spikes, and more rapid prespike trajectory (from −179 ms to threshold) with CRF + nCRF stimulation compared with CRF alone stimulation (dV/dt _CRF = 0.062 ± 0.002 mV/ms; dV/dt _{CRF + nCRF} = 0.073 ± 0.002 mV/ms; p < 0.01). Traces aligned at spike threshold voltage before averaging (0 on ordinate). Inset shows that spike threshold is also significantly lower with wide-field stimulation (Threshold _{CRF + nCRF} = −55.1 ± 0.2 mV; Threshold _CRF = −54.2 ± 0.2 mV; p < 0.01). All data for n = 13 RS_C neurons (mean ± SEM).

Is increased spike time precision across trials accompanied by a decrease in the width of the synaptic barrages triggering spikes? We calculated the spike-triggered average (STA) of V_m across the population of RS_C neurons (n = 13) in response to CRF and CRF + nCRF stimulation. Membrane potential STAs were systematically sharper in the CRF + nCRF condition (Figure 5B; red versus black), and accompanied by an increase in the average rate of change of V_m (dV/dt) for the rising phase of V_m trajectory to spike threshold (Figure 5B). This change does not appear to result from a hyperpolarization-induced increase in driving force on EPSPs, because DC-induced hyperpolarization of RS_C neurons during CRF alone presentation did not result in the same degree of STA sharpening seen during CRF + nCRF stimulation (n = 10 RS_C neurons; data not shown).

Consistent with the finding of a more hyperpolarized membrane potential prior to spike initiation and a faster prespike dV/dt, actual spike thresholds (defined as the voltage at the peak of the second derivative of V_m) were significantly lower across the population of RS_C neurons during CRF + nCRF stimulation compared with CRF-alone stimulation (Figure 5B; see also Azouz and Gray, 2003).

Interaction of Synaptic Excitation and Inhibition Largely Explains Changes in Spike Responses during CRF + nCRF Stimulation

Our experimental results demonstrate that in RS_C neurons IPSPs become stronger without changes in reliability during CRF + nCRF stimulation, while conversely, EPSPs become more reliable with no change in average amplitude. How do these two factors contribute to changes in spike train sparseness and reliability?

Experimental isolation of EPSPs and IPSPs necessitates pharmacological blockade of intrinsic conductances, making it impossible to simultaneously record EPSPs, IPSPs, and spikes, thereby preventing us from directly addressing this question (but see Pospischil et al., 2007). Instead we turned to a simple leaky integrate and fire (LIF) single-neuron model where we could explore the relative contributions of recorded EPSPs and IPSPs on the sparseness and reliability of spiking activity. Excitatory and inhibitory conductances (G_e and G_i) were derived from our recordings of isolated EPSPs and IPSPs, respectively, and then simultaneously injected into the model LIF neuron at rest (−65 mV; e.g., Figure 6C; see Supplemental Experimental Procedures).

Figure 6. Changes in Excitatory and Inhibitory Synaptic Barrages Drive Increased Sparseness and Reliability with Wide-Field Stimulation in a Leaky Integrate and Fire Model Neuron.

(A) Correction for input resistance (R_in) and capacitance (C_m) of the recorded neuron allows inference of synaptic currents (IPSC or EPSC) that underlie an individual IPSP (left) or EPSP (right) amplitude-time series recorded in real neurons during CRF presentation. All traces in this figure were derived from data obtained from the neuron illustrated in Figures 1 and 2.

(B) Injection of these IPSC or EPSC traces into a leaky integrate and fire (LIF) model with experimentally measured R_in and C_m reproduces the original recorded IPSP (left) and EPSP (right) trace. Reconstructed example EPSP and IPSP amplitude-time series for CRF + nCRF stimulation shown in blue and red (lower traces).

(C) Excitatory and inhibitory conductances (G_e and G_i) derived from the reconstructed currents during CRF stimulation are injected into the LIF model cell at rest (−65 mV).

(D) Matrix of G_e and G_i combinations that can be examined in the LIF model. Injection of G_e and G_i from the same conditions (e.g., within CRF or CRF + nCRF stimulation) represents the control conditions (Da and Db). Mixing G_e and G_i obtained from different conditions represents our experimental manipulation (Dc and Dd).

(Ea) LIF raster and PSTH in response to 60 simulated (E + I)_CRF trials. Sparseness, S = 0.32 ± 0.002, spike-train reliability, R_CRF = 0.33 ± 0.02. (Eb) LIF raster and PSTH in response to 60 simulated (E + I)_{CRF + nCRF} trials. Sparseness and spike-train reliability increase significantly (S = 0.70 ± 0.002, spike train R_{CRF + nCRF} = 0.41 ± 0.02; p < 0.01 for both comparisons to CRF simulations). Note nonlinear change in shape of PSTH: some peaks are enhanced (solid arrowheads) while others are suppressed (open arrowhead). Correlation coefficient (r) of PSTH (E + I)_CRF to PSTH (E + I)_{CRF + nCRF} = 0.46 ± 0.02. (Ec) LIF raster and PSTH in response to 60 simulated (E_CRF + I_{CRF + nCRF}) trials. Sparseness increase significantly (S = 0.68 ± 0.002; p < 0.01) compared with (E + I)_CRF. Spike-train reliability decreased significantly in comparison to (E + I)_CRF simulations. (E_CRF + I_{CRF + nCRF}) R = 0.33 ± 0.02; p < 0.01. Correlation coefficient (r) of PSTH (E_CRF + I_{CRF + nCRF}) to PSTH (E + I)_{CRF + nCRF} = 0.47 ± 0.02. (Ed) LIF raster and PSTH in response to 60 simulated (E_{CRF + nCRF} + I_CRF) trials. Sparseness increased (S = 0.53 ± 0.002; p < 0.01), although significantly less than in (E_CRF + I_{CRF + nCRF}) simulation. However, spike-train reliability increased significantly in comparison to (E_CRF + I_{CRF + nCRF}) simulation (p < 0.01), and was not significantly different from (E + I)_CRF simulations, (E_{CRF + nCRF} + I_CRF) R = 0.40 ± 0.02; p > 0.1. Correlation coefficient of PSTH (E_{CRF + nCRF} + I_CRF) to PSTH (E + I)_{CRF + nCRF} = 0.97 ± 0.02, a significant (106%, p < 0.01) increase compared with (E_{CRF + nCRF} + I_CRF) simulations.

How well does combined G_e and G_i injection into the LIF model replicate a real neuron’s spiking response? We examined this by utilizing spiking responses recorded from an RS_C neuron prior to onset of action potential inactivation (Figure 1). G_e and G_i conductance traces were derived for this same neuron based on responses to the same movie, but after blockade of intrinsic conductances, as described above (Figure 2). G_e and G_i were derived by compensating for the resistive-capacitive properties of the recorded neuron at the actual holding potentials used to record EPSPs and IPSPs (Figure 6A; see Nowak et al., 1997). The full amplitude-time series of the underlying conductance (G_e or G_i) was then calculated from the amplitudes of the reconstructed synaptic currents and the instantaneous membrane potential driving force. Each derived G_e and G_i time series was then injected into an LIF model neuron matched in input resistance and membrane time constant to the recorded neuron. The model neuron’s membrane potential accurately replicated the original IPSP and EPSP barrages, when injected at the same holding potentials (cf. Figures 6A and 6B, top traces), confirming the model’s basic validity. The same procedure was applied to all single trial EPSP and IPSP responses in the CRF and CRF + nCRF conditions, and used to construct a database of G_e and G_i traces for CRF and CRF + nCRF stimulation. Individual G_e and G_i conductance traces were drawn at random from the database (within a given stimulus condition), and injected simultaneously into the model neuron at rest to generate simulated single trial intracellular membrane potential and spike responses (Figure 6C). This procedure was repeated for 60 unique combinations of G_e and G_i sequences in each stimulus configuration.

As seen in Figure 6Ea, random combinations of G_e and G_i sequences recorded during CRF-alone stimulation produce simulated spike responses that are highly correlated with the actual recorded spike response for the same neuron during CRF stimulation (cf. Figure 1C; Model PSTH versus Actual PSTH r_CRF = 0.54 ± 0.006). The model neuron, like the real neuron, showed a nonlinear change in the PSTH in response to (E + I)_{CRF + nCRF} injection (Figure 6Eb, arrows) and a dramatic increase in response sparseness (compare Figures 6Ea and 6Eb and Figures 1C and 1D). Response sparseness of the model neuron increased significantly during “wide-field” stimulation; similarly, the trial-to-trial reliability of the spike trains increased significantly in the (E + I)_{CRF + nCRF} simulation. These computational results strongly suggest that increased spike train sparseness and reliability (at rest) are largely accounted for by simply combining the underlying excitatory and inhibitory conductances evoked by CRF + nCRF stimulation.

Simulations Support the Hypothesis that Network Inhibition Drives Changes in Sparseness while Recurrent Excitation Drives Changes in Reliability

The model was then used to examine the relative contributions of excitation and inhibition upon sparseness and reliability during CRF + nCRF stimulation. This was done by artificially pairing G_e traces derived from CRF recordings with G_i traces derived from CRF + nCRF recordings (Figure 6Dc; E_CRF + I_{CRF + nCRF}), and vice versa. As evident in the PSTH, adding I_{CRF + nCRF} trials to E_CRF trials significantly increased sparseness compared with pairing of I_CRF with E_CRF trials, but did not significantly alter spike-train reliability (Figure 6Ec).

However, when we paired G_e derived from CRF + nCRF presentation with G_i derived from CRF recordings, we observed a dramatic increase in the similarity of the PSTH to that constructed from the (E + I)_{CRF + nCRF} trials, along with a significant increase in the trial-to-trial reliability of the spiking responses that was no different than the reliability of the spike trains in the full (E + I)_{CRF + nCRF} simulation. Conversely, the increase in sparseness was significantly smaller than that observed in the E_CRF + I_{CRF + nCRF} simulation. The results from this neuron suggest that changes in G_i have a predominant effect on sparseness, while changes in G_e have a predominant effect on spike reliability and the similarity of PSTH structure to that occurring normally under wide-field stimulation conditions.

We repeated these G_e and G_i LIF simulations for all of the neurons from which we recorded EPSPs and IPSPs (n = 9; same as in Figure 2C), and the simulation results strongly paralleled many of our experimental observations. First, in every case individually, and across the population, coinjection of G_e and G_i derived from CRF + nCRF stimulation produced significantly sparser (34%) and more reliable (27%) spike trains in comparison to injection of G_e and G_i derived from CRF stimulation alone (Figures 7A and 7B).

Figure 7. LIF Simulations Indicate that Network-Generated Inhibition Drives Increased Sparseness, while Network-Generated Excitation Drives Increased Response Reliability.

(A) LIF simulations derived from population of real neurons (n = 9; see Figure 2) indicate that spike-train sparseness was significantly lower in the (E + I)_{CRF + nCRF} simulations compared with the three other simulation conditions (S = 0.47 ± 0.04; colored asterisks indicate significant group differences corrected for multiple comparisons). Response sparseness in the E_CRF + I_{CRF + nCRF} simulations (blue; S = 0.65 ± 0.04) was not significantly different than the (E + I)_{CRF + nCRF} simulation (violet; S = 0.65 ± 0.03; p > 0.1), but both of these groups displayed significantly larger response sparseness than the E_{CRF + nCRF} + I_CRF simulations (red; S = 0.58 ± 0.03; p < 0.01 for both group comparisons). Colored asterisks indicate significant group differences. Values are mean ± SEM.

(B) Conversely, trial-to-trial spike-train reliability is highest for the (E + I)_{CRF + nCRF} simulations (violet; R = 0.2 ± 0.007), and these spike trains were not significantly more reliable than those in the E_{CRF + nCRF} + I_CRF simulations (red; R = 0.19 ± 0.007; p > 0.1). However, spike responses of both of these groups were significantly more reliable than the spike trains of the E_CRF + I_{CRF + nCRF} simulations (blue; R = 0.17 ± 0.006; p < 0.01 for both group comparisons).

(C) Summary of LIF simulations shows that overall spiking pattern (PSTH) of E_{CRF + nCRF} + I_CRF simulations is the most similar to (E + I)_{CRF + nCRF} PSTH (red; r = 0.74 ± 0.007), although the E_CRF + I_{CRF + nCRF} PSTH was significantly more similar to the (E + I)_{CRF + nCRF} PSTH (blue; r = 0.46 ± 0.008), as compared with the similarity of the (E + I)_CRF PSTH to the (E + I)_{CRF + nCRF} PSTH (black; r = 0.26 ± 0.007; p < 0.01 for all group comparisons). See also Figure S6 for effects of manipulating timing of G_i relative to G_e.

Second, spikes occurring in the CRF + nCRF simulations were accompanied by significant narrowing of the width of the average synaptic conductance preceding each spike. Furthermore, jittering the exact timing of the excitatory-inhibitory relationship by as little as 20 to 50 ms significantly decreased spike-train reliability and sparseness, respectively (Figure S6).

Most importantly, by artificially recombining the influences of CRF versus CRF + nCRF induced excitation and inhibition, we found that changes in G_i during CRF + nCRF stimulation had a predominant effect on increasing sparseness (Figure 7A), while changes in G_e during CRF + nCRF stimulation had a predominant effect on increasing spike-train reliability (Figure 7B). Moreover, changes in G_e largely dictated the shape of the overall PSTHs compared with those obtained with (E + I)_{CRF + nCRF} stimulation (Figure 7C). These results indicate that the synchronous interaction of G_e and G_i induced by combined CRF + nCRF stimulation largely replicates the effects observed in our recordings, with enhanced inhibition contributing to increased sparseness, which in turn facilitates more reliable and precise recurrent cortical excitation that determines the overall spiking response.

DISCUSSION

We have demonstrated here that wide-field naturalistic visual stimulation—simultaneously engaging the CRF + nCRF—not only increases response sparseness in cat primary visual cortex, but also significantly increases trial-to-trial reliability and temporal precision. These effects arise as a consequence of complex interactions between excitatory and inhibitory mechanisms mediated by distinct neuronal subtypes. Wide-field visual stimulation simultaneously increased activity of FS inhibitory interneurons and increased IPSP amplitudes in RS_C pyramidal neurons. At the same time, in the same population of RS_C neurons, wide-field stimulation produced an increase in the trial-to-trial reliability of both action potentials and underlying EPSPs. Interestingly, the injection of excitatory and inhibitory conductances (derived from actual recordings of RS_C neuron responses) into a simple model replicated these findings. This suggests that changes in visually evoked synaptic potentials during wide-field stimulation were largely responsible for the observed increases in action potential sparseness and reliability in RS_C neurons. The simulations also revealed that increased amplitudes of visually evoked IPSPs predominantly drove increased neuronal sparseness, while increased EPSP reliability predominantly drove increased action potential reliability.

Functional Interactions between Inhibitory and Excitatory Networks Determines Response Sparseness and Reliability

Cortical neuronal responses are determined in large part by the precise amplitude-time course of barrages of excitatory and inhibitory synaptic potentials (Haider and McCormick, 2009; Pouille and Scanziani, 2001; Tiesinga et al., 2008) interacting with the intrinsic membrane properties of the neuron. Our LIF simulations (necessarily absent of intrinsic conductances) utilizing naturally occurring synaptic responses were able to reproduce changes in sparseness and reliability as observed in real RS_C neurons. This suggests that the wide-field induced changes in EPSPs and IPSPs, coupled with the nonlinear dynamics of spike generation, are nearly sufficient to explain increased spike-train sparseness and reliability during naturalistic sensory stimulation.

At the level of synaptic potentials, one of the key mechanisms contributing to increased sparseness of RS_C pyramidal neurons appears to be activation of cortical inhibitory networks, resulting in an increase in the amplitude, and modulation of the timing, of IPSP barrages. One component of enhanced inhibition that may contribute to increased sparseness in RS_C neurons is an “iceberg” effect (see Carandini, 2007 and references therein), in which responses to weak synaptic inputs become sub-threshold, while the strong inputs (which may be less numerous than the weak ones) remain suprathreshold, resulting in an increase in overall response sparseness. Although tonic hyper-polarization via current injection does increase sparseness, this increase is smaller (per mV of hyperpolarization) than that associated with natural IPSP barrages generated during wide-field stimulation. These results suggests that the precise timing of wide-field activated IPSP barrages relative to the EPSPs evoked by the same stimuli is an important factor in determining the overall structure of synaptic potentials that shape neuronal selectivity.

Supporting this hypothesis is the observation that the spike triggered average membrane potential in RS_C cells is sharper and rises more rapidly during wide-field stimulation (Figure 5B). Accordingly, our simulations demonstrate that the initiation of action potentials in the wide-field condition is, on average, associated with a more rapid increase in synaptic excitation with a simultaneous decrease in inhibition (Figure S6; Hasenstaub et al., 2005). Disrupting this temporal relationship between wide-field elicited EPSP and IPSP barrages by tens of milliseconds significantly decreases sparseness and reliability in simulations of CRF + nCRF spiking (Figure S6). These effects reflect the importance of the average temporal relationship of excitation and inhibition across trials; the timing and covariance of excitation with inhibition is likely even more precise within a single trial (Okun and Lampl, 2008).

Our simulations revealed that the changes in IPSPs elicited by wide-field stimulation have a larger effect on sparseness than the changes occurring in EPSPs (Figure 7). Conversely, changes in EPSPs occurring with wide-field stimulation had a significantly stronger effect on spike train reliability than the IPSPs. We propose a simple scenario that may underlie the generation of sparse yet reliable sensory responses during wide-field stimulation (Figure 8). The increased activation of inhibitory neurons results in an overall decrease in action potential responses to weak excitatory synaptic inputs to RS_C cells. This results in an increase in neuronal sparseness in individual RS_C neurons. Increased reliability emerges because action potentials that do occur in RS_C neurons happen preferentially during the peaks of sharply depolarizing synaptic potentials, which are restricted to narrower windows of time, resulting in less temporal jitter and greater reliability in the generation of spikes (Rodriguez-Molina et al., 2007). This increased spiking reliability and precision in RS_C neurons is communicated as reliable and sharper EPSPs to postsynaptic neurons, many of which then further increase spike reliability (and decrease temporal jitter) through a similar mechanism. Reverberation or passing through multiple stages of the recurrent cortical network could then result in strong temporal sharpening of neural responses and increase the reliability and selectivity of spiking across the population (Litvak et al., 2003; Wang et al., 2006).

Figure 8. Schematic Diagram of Proposed Excitatory-Inhibitory Interactions during Wide-Field Visual Stimulation.

(A) Local cortical networks composed of excitatory (white) and inhibitory (black) neurons form interconnections with each other, with the great majority of connectivity occurring among excitatory neurons. During CRF stimulation, both excitatory and inhibitory cell types are driven, with RS_C neurons and FS neurons generating elevated and temporally varying responses (traces).

(B) Upon simultaneous engagement of the CRF and nCRF, inhibitory interneurons become strongly activated by increased excitatory drive arising from a larger spatial distribution of inputs. The increased depolarization and enhanced synaptic fluctuations in interneurons are nonlinearly transformed into greater numbers of spikes compared with excitatory neurons (inset at center). This causes RS_C neurons to receive enhanced inhibitory synaptic barrages at specific time points, which leads to increased sparseness and precision of visually evoked spike responses in RS_C neurons. These sparser but less variable spikes are amplified through the recurrent excitatory connections among RS_C neurons in the local network (red synapse), which leads to more reliable and precise sensory encoding across the ensemble of pyramidal neurons (red trace).

Mechanisms Underlying Increased Activity in Inhibitory Circuits during Wide-Field Stimulation

The data presented here strongly support the idea that increased IPSP amplitudes are a critical factor underlying increased sparseness. Three changes that can increase the size of IPSP barrages are: (1) An increase in the number, or intensity, of discharging excitatory neurons presynaptic to the inhibitory neurons; (2) an increase in the synchrony of activity within the network of presynaptic excitatory neurons; and (3) a direct increase in the activity level of the GABAergic inhibitory neurons due to nonlinear response properties.

We recorded from two commonly encountered subtypes of excitatory pyramidal cells: classical regular spiking (RS_C) and thin-spike regular spiking (RS_TS) neurons (Nowak et al., 2003). RS_C pyramidal neurons, thought to be the most common cortical subtype, decreased their average spike responses to wide-field visual stimulation. By itself, this finding would suggest a reduced excitatory drive of both pyramidal neurons and local GABAergic neurons.

However, several key observations are at odds with this inference. First, even though the overall firing rate of RS_C neurons decreased during wide-field stimulation, the response to a subset of stimulus frames was often accompanied by increased peak firing rates (e.g., Figures S1, S5) and was more precise (Figure 5) and repeatable (Figure 4). This concerted increase in temporal precision and population reliability could well facilitate intracortical synchrony, an especially effective driver of FS and non-FS GABAergic neurons (Jonas et al., 2004; Kapfer et al., 2007). Second, RS_TS pyramidal neurons exhibited significantly increased mean and peak firing rates during wide-field stimulation, perhaps contributing preferentially to enhanced FS neuronal activity. Third, CRF + nCRF stimulation necessarily activates a larger cortical area and therefore increases the total number of active excitatory neurons in the local cortical network. Many of these cells may be presynaptic to local GABAergic neurons, or activate neurons that are so (e.g., RS_TS pyramidal cells). Although we could only record from one subtype (FS) of cortical interneuron, we did find that their receptive fields were larger than those of RS_C neurons, suggesting that at least some GABAergic neurons may preferentially receive excitatory inputs from wider regions of sensory space (Bruno and Simons, 2002; Liu et al., 2009; Wu et al., 2008). These findings along with anatomical observations suggest that the IPSPs we recorded from RS_C neurons during wide-field stimulation likely resulted from the activation of local (i.e., within 2.5 mm of the recorded neuron) inhibitory inter-neurons (Kisvarday and Eysel, 1993), which are themselves driven by a diverse set of local and long-range excitatory connections (Gilbert et al., 1996; Martin and Whitteridge, 1984; McGuire et al., 1991) originating from both the CRF and nCRF (Figure 8).

Perhaps the most important factor contributing to increased inhibitory neuron activity is the relationship between membrane potential and firing rate. In the presence of noisy synaptic activity, this function takes the form of a power law: f = (V)^x (Miller and Troyer, 2002). We found that FS interneurons exhibit a particularly steep power law relationship, with a relatively large exponent compared with RS_C neurons (Figure S3). As a result, FS cells are likely to be more sensitive to and strongly activated by the overall synaptic fluctuations provided by joint stimulation of the CRF and nCRF (Figure 8).

One should keep in mind that although we only recorded from FS interneurons, the effects of inhibition on RS_C neurons arise from a broad range of inhibitory neuron subtypes. Indeed, other subtypes of interneurons also generate highly nonlinear response enhancement to costimulation of multiple excitatory pathways (e.g., Martinotti cells; Kapfer et al., 2007; Silberberg and Markram, 2007). Clarification of the specific contributions of the many inhibitory neuron subtypes to response selectivity and reliability requires further investigation.

Implications of Increased Sparseness, Reliability, and Precision

One of the main observations of our study was that classical regular spiking pyramidal neurons (RS_C) simultaneously increase their visual selectivity and reliability, while decreasing overall firing rate, in response to wide-field visual stimulation. Similarly, presentation of full field natural scenes in a temporal sequence that replicates natural eye movements also generated synaptic and action potential responses that are both reliable and sparse in cat V1 (Yves Fregnac, Pierre Baudot, Manuel Levy, and Olivier Marre, 2005, Cosyne meeting, abstract). This finding suggests that, during natural vision, RS_C neurons are more energetically efficient (Niven and Laughlin, 2008), more selective (Figure 1D) and reliable (Figure 4C), and consequently more informative per spike (Vinje and Gallant, 2002). Our observations provide strong support for the efficient sparse coding hypothesis (Barlow, 1972; Olshausen and Field, 2004; Vinje and Gallant, 2000, 2002).

The moment-to-moment demands of natural behavior depend upon the reliability of neuronal responsiveness, and insights into mechanisms that limit response variability are critical toward understanding the nature of cortical computation. Indeed, repeated presentations of identical stimuli can elicit highly variable spike responses (Heggelund and Albus, 1978; Shadlen and Newsome, 1998; Tiesinga et al., 2008). Although some studies have shown that response reliability can be quite high after controlling for factors such as eye movements or recording mainly from input layers (Gur et al., 1997; Kara et al., 2000), in general, the variability of cortical responses scales approximately with the mean (but see DeWeese and Zador, 2006; Maimon and Assad, 2009). These conclusions are based on recordings likely to be dominated by RS_C neurons, because they are by far the most commonly recorded cells in visual cortex (but see Chen et al., 2008; Mitchell et al., 2007).

To our knowledge, the data presented here provide the first mechanistic link between sparse sensory coding and increased response reliability in visual cortex under naturalistic stimulus conditions. These findings complement recent findings of sparse cortical responses in several other species and sensory systems, under a variety of experimental conditions (Greenberg et al., 2008; Houweling and Brecht, 2008; Hromadka et al., 2008; Waters and Helmchen, 2006). Although recent in vitro studies have shown that response correlation necessarily increases as firing rates increase (de la Rocha et al., 2007), we show here that sparse sensory responses, which exhibit increased efficiency (i.e., fewer total action potentials), can be accompanied by increased response reliability across trials. These findings have broad implications for the nature of visual encoding, since they demonstrate that increased sparseness (reflective of increased stimulus selectivity) is also associated with higher trial-to-trial response reliability to a more restricted set of stimuli.

Relationship to Previous Studies and Other Sensory Systems

We have identified a basic operational mode of visual cortical circuits engaged by spatiotemporally rich wide-field stimulation, as experienced during natural vision (David and Gallant, 2005; Mazer and Gallant, 2003). Our work suggests that intracortical inhibitory networks are critical for the generation of selective and reliable visual responses during wide-field stimulation. Specifically, we hypothesize that the intrinsic properties and anatomical connectivity of FS and other types of interneurons, as well as the RS_TS subclass of pyramidal neuron, enable them to rapidly integrate activity from horizontal connections and shape the output of RS_C pyramidal cells. We speculate that an additional function of the excitatory-inhibitory network mechanisms described here is to decorrelate neuronal responses across cell classes (David and Gallant, 2005; Mazer and Gallant, 2003; Vinje and Gallant, 2000; Wang et al., 2003), ultimately increasing coding bandwidth and efficiency (El Boustani et al., 2009; Felsen et al., 2005; Olshausen and Field, 1996; Simoncelli, 2003; Vinje and Gallant, 2002), while simultaneously increasing signal reliability across the population.

Our results are generally consistent with the large body of extracellular recording literature indicating an overall suppression of CRF elicited responses with nCRF costimulation (Ange-lucci and Bressloff, 2006; Bair et al., 2003; Cavanaugh et al., 2002b; DeAngelis et al., 1994; Durand et al., 2007; Fitzpatrick, 2000; Jones et al., 2001; Webb et al., 2005), and increased spiking selectivity driven by surround stimulation (Chen et al., 2005; Okamoto et al., 2009). Our observations are also consistent with studies showing that stimulus context modulates both perceptual and neuronal sensitivity (Ito and Gilbert, 1999). However, our results extend these findings by demonstrating that wide-field visual stimulation with dynamic, spatiotemporally rich stimuli drives highly specific network interactions among distinct neuronal subtypes that ultimately lead to increases in spike precision, spike reliability, and response sparseness in the output of RS_C pyramidal neurons.

Moreover, the results presented here during naturalistic stimulation strongly implicate intracortical inhibitory potentials, at least partially originating in FS interneurons, as a critical component of these effects (cf. Anderson et al., 2001; Ozeki et al., 2009). Although complex inhibitory modulations are not entirely unexpected from a spatiotemporally rich and time-varying stimulus, the network interactions described here among distinct cortical neuronal subtypes that ultimately increase the reliability and precision of the network response are not easily predicted from existing studies of surround suppression. It is entirely possible that the use of a spatiotemporally rich wide-field (“naturalistic”) stimulus set puts the cortex in a more “transient” response regime that strongly engages inhibitory circuits (Ozeki et al., 2009). Such dynamics certainly deserve further investigation.

Finally, it is likely that the excitatory-inhibitory mechanisms described here in visual cortex generalize across sensory systems, particularly under naturalistic stimulus conditions. Recent studies of the rodent somatosensory system show that cortical responses exhibit nonlinear modulation with multi-whisker stimulation as compared with single whisker stimulation, and elicit complex activity patterns across extended regions of barrel cortex (Jacob et al., 2008). Studies of both mammalian auditory cortex (Hromadka et al., 2008) and the avian song system (The-unissen et al., 2001) indicate that neural responses are highly selective for features present in natural sounds. Song generation itself may be mediated by “ultra sparse” neuronal activity (Hahn-loser et al., 2002). Although our results demonstrate the intracellular and network mechanisms underlying enhanced selectivity to ongoing wide-field visual stimulation, the exact interaction of the spatiotemporal statistics of natural sensory stimulation with the response properties of distinct neuronal subtypes merits further investigation.

EXPERIMENTAL PROCEDURES

Animal Preparation and Electrophysiological Recordings

Briefly, young adult female cats were initially anesthetized with ketamine/xylazine and then maintained on isoflurane vaporized in O₂ for the duration of the experiment. Standard surgical procedures for reducing respiratory and cardiac pulsations were employed, and all experiments conformed to Yale University IACUC standards. A craniotomy overlying Area 17 was performed, the dura was dissected, and metal electrodes and/or beveled sharp glass micropipettes (55–120 MΩ), filled with 2 M K⁺ acetate (for recording action potentials) or filled with 25–50 μM Qx-314 and 2 M Cs⁺ acetate (to block most intrinsic conductances; see Haider et al., 2006; Hasenstaub et al., 2005) were advanced into the cortex. All action potential responses to natural movies were recorded with zero current injection; EPSPs were recorded near −80 mV and IPSPs were recorded near 0 mV.

Visual Stimulation

Neurons were characterized with computer assisted hand-mapping, then quantitatively mapped using a two-dimensional (2D) sparse noise stimulus (Jones and Palmer, 1987; Mazer et al., 2002) composed of light and dark bars at each neuron’s preferred orientation on a linearized 19 inch CRT (Siemens). Screen background was a uniform gray and bars were 100% contrast. We designated the least-squares 2D circular fit of the half-maximal spike response contour (or the half-maximal membrane potential response in the experiments where spikes were inactivated) as the CRF. We then presented repeated segments of one of seven different long-duration (5–16 s; without jumps or cuts) movies for 10–20 trials, and determined which frame of the movie evoked the greatest number of spikes. We then selected this frame and 1.5 s flanking each side of this frame to present as the 3 s “optimal” movie, as shown in all figures here. A mask of equal color and luminance with the background occluded all portions of the movie save for a circle of diameter equal to and centered over the CRF. This mask was enlarged to expose 3X the CRF, and these stimuli were designated as the CRF + nCRF stimuli. In both cases, the pixels presented in the CRF were identical over trials. Presentation of CRF alone and CRF + nCRF stimuli were randomly interleaved. Movies were digitized from commercial DVDs (Winged Migration, The Incredibles, Aeon Flux), converted to grayscale and presented at 25–28 Hz.

Analysis

We quantified neural selectivity by computing lifetime response sparseness (S), S = 1 – a, where a denotes the activity fraction, a = [Σ_i (r_i/n)]²/Σ_i (r_i²/n), and r_i is the response to the i-th frame of the movie, and n is the total number of frames in the movie. Lifetime sparseness (S) is a metric of a single neuron’s selectivity that is closely related to the kurtosis of the firing rate distribution. For highly selective neurons, with maximal responses occurring primarily during a single movie frame, the response distribution across all movie frames will be highly peaked and S will approach 1.0 (Willmore and Tolhurst, 2001). Lifetime sparseness is different from population sparseness, which measures the activity profile across an ensemble of neurons.

Real-time analysis and visual stimulation utilized custom written software in Python (PyPE). All arithmetic means reported and plotted ± standard error of the mean. All analyses, statistics and plots were generated with built-in and custom functions in MATLAB (Mathworks). Unless explicitly noted, all p values (α = 0.01) were calculated with the nonparametric Kruskal-Wallis analysis of variance, with Tukey-Kramer correction in cases of multiple comparisons. Throughout the main text and results, when statistical significance is mentioned, p values and tests are presented in the appropriate portions of the corresponding figure legends or tables, as relevant.