The Dynamical Regime of Sensory Cortex: Stable Dynamics around a Single Stimulus-Tuned Attractor Account for Patterns of Noise Variability

Abstract

Correlated variability in cortical activity is ubiquitously quenched following stimulus onset, in a stimulus-dependent manner. These modulations have been attributed to circuit dynamics involving either multiple stable states ("attractors") or chaotic activity. Here we show that a qualitatively different dynamical regime, involving fluctuations about a single, stimulus-driven attractor in a loosely balanced excitatory-inhibitory network (the stochastic "stabilized supralinear network"), best explains these modulations. Given the supralinear input/output functions of cortical neurons, increased stimulus drive strengthens effective network connectivity. This shifts the balance from interactions that amplify variability to suppressive inhibitory feedback, quenching correlated variability around more strongly driven steady states. Comparing to previously published and original data analyses, we show that this mechanism, unlike previous proposals, uniquely accounts for the spatial patterns and fast temporal dynamics of variability suppression. Specifying the cortical operating regime is key to understanding the computations underlying perception.

Keywords: MT; V1; circuit dynamics; cortical variability; noise correlations; theoretical neuroscience; variability quenching.

Figure 1

Three Different Dynamical Regimes that Could Explain Variability Modulation by Stimuli (A–C) Two schematic neural trajectories (red and green) corresponding to two separate trials are plotted for each dynamical regime, before (top) and after (bottom) stimulus onset. Spontaneous activity is redrawn in gray beneath evoked activity to allow comparison of variability. Dotted ellipses outline activity covariances around the fixed point(s) of the dynamics (if any exist). (A) Multi-attractor dynamics: spontaneous activity wanders stochastically between a set of attractor states (three shown), resulting in large trial-by-trial variability (top). Stimulus onset constrains fluctuations to the vicinity of a single attractor, reducing variability across both time and trials (bottom). (B) Chaos suppression: chaos yields large across-trial variability in spontaneous dynamics (top), which is suppressed by the stimulus, leading to a reduction of variability across trials but not necessarily across time (bottom). (C) Stochastic SSN: both spontaneous and evoked dynamics are stable with a single fixed point, but the stimulus can shrink the effective size of the basin of attraction of the fixed point (as well as shifting its location), resulting in a reduction of both across-time and across-trial variability.

Figure 2

Activity Variability in a Reduced, Two-Population Stochastic SSN (A) The network is composed of two recurrently connected units, summarizing the activity of two populations of excitatory (red) and inhibitory (blue) neurons. Both units receive private input noise and a common constant input h. (B) Threshold-quadratic neural input/output function determining the relationship between membrane potential and momentary firing rate of model neurons (Equation 2). (C) Sample $V_{\mathrm{E}/\mathrm{I}}$ traces for the two units (top), as the input is increased in steps from $h=0$ to 2 mV to 15 mV (bottom). (D) Dependence of population activity statistics on stimulus strength h. Top: mean E (red) and I (blue) firing rates; middle: mean $V_{\mathrm{E}/\mathrm{I}}$; bottom: standard deviation of $V_{\mathrm{E}/\mathrm{I}}$ fluctuations. The comparison with a purely feedforward network $\left(\mathbf{W}=\mathbf{0}\right)$ receiving the same input h is shown in gray. Dots are based on numerical simulations of 500 trials. Solid lines show analytical approximations (Hennequin and Lengyel, 2016).

Figure 3

Modulation of Variability in a Randomly Connected Stochastic Spiking SSN (A) Top: raster plot of spiking activity, for 40 (out of 4,000) excitatory neurons (red) and 10 (out of 1,000) inhibitory neurons (blue). Upper middle: momentary E and I population firing rates. Lower middle: LFP (momentary population-averaged $V_{\text{m}}$). Bottom: $V_{\text{m}}$ of two randomly chosen excitatory neurons. The dashed vertical line marks the onset of stimulus, when h switches from 2 mV to 15 mV. Population firing rates, LFP, and $V_{\text{m}}$ traces were smoothed with a Gaussian kernel of 50 ms width. (B) Top, normalized LFP power in spontaneous (black) and evoked (green) conditions; bottom, average ($\pm$ SEM) spectral coherence between single-cell $V_{\text{m}}$ and the LFP; left, model; right, data from V1 of the awake monkey, reproduced from Tan et al. (2014).

Figure 4

Modulation of Variability in a Stochastic SSN with a Ring Architecture (A) Schematics of the ring architecture. Excitatory (red) and inhibitory neurons (blue) are arranged on a ring, their angular position indicating their preferred stimulus (expressed here as preferred stimulus orientation, PO). The stimulus is presented at $0^{\circ }$. (B) Synaptic connectivities all follow the same circular Gaussian profiles with peak strengths that depend on the type of pre- and post-synaptic populations (excitatory, E, or inhibitory, I). (C) Each neuron receives a constant input with a baseline (black line, 0% contrast), which drives spontaneous activity, and a tuned component with a bell-shaped dependence on the neuron’s preferred orientation and proportional to contrast, c (dark and light green, 50% and 100% contrast, respectively). Neurons also receive spatially and temporally correlated noise, with spatial correlations that decrease with tuning difference (see Figure 5D). (D) Single-trial network activity (E cells), before and after the onset of the stimulus (100% contrast). Neurons are arranged on the y axis according to their preferred stimuli. (E) Reduction in membrane potential variability across trials: membrane potential traces in 5 independent trials (top) and Fano factors (bottom) for an E cell tuned to the stimulus orientation (left) or tuned to the orthogonal orientation (right). For $V_m$, orange and brown lines and shading show (analytical approximation of) across-trial mean $\pm$ SD. (F) Reduction of average spike count Fano factor in the population following stimulus onset in the model (top) and experimental data (bottom). Spikes were counted in 100 ms time windows centered on the corresponding time points. (G) Mean firing rates (top), Fano factors (middle), and std. of voltage fluctuations (bottom) at different contrast levels as a function of the neuron’s preferred stimulus in the model (left) and, for rate and Fano factor, experimental data (right, averaged across 99 neurons). Colors indicate different contrast levels (model: colors as in C; data: black, spontaneous, green, 100% contrast). (H) Shared variability in normalized spike counts, as estimated via factor analysis (STAR Methods; Churchland et al., 2010), before (spontaneous, black) and after stimulus onset (evoked, green) in the model (left) and experimental data (right). Dots in (F) and (G) are based on numerical simulations of 500 trials. For the model, colored lines and shaded areas in (E) and solid lines in (F) and (G) show analytical approximations (Hennequin and Lengyel, 2016). Experimental data analyzed in (F)–(H) are from awake monkey V1 (Ecker et al., 2010), with error bars denoting 95% CI.

Figure 5

Low-Dimensional Bump Kinetics Explain Noise Variability in the Ring SSN (A) Sample of $V_{\text{m}}$ fluctuations across the network in the evoked condition (left, “true activity,” 100% contrast), to which we fitted a circular-Gaussian function (bump) $V_i\left(t\right)=a\left(t\right)\text{exp}\left[\left(cos\left({\theta }_i-\mu \left(t\right)\right)-1\right)/{\sigma }^2\left(t\right)\right]$ across the excitatory population in each time step (center), parametrized by its location, μ, and width, σ. The amplitude of the bump, a, was chosen in each time step so as to keep total population firing rate constant. Fluctuations in location and width were independent, and the fit captured $87\text{\%}$ of the variability in $V_{\text{m}}$ (right). (B) The two principal modes of bump kinetics: small changes (red arrows) in location (top) and width (bottom) of the activity bump result in the hill of network activity deviating from the prototypical bump (gray shadings). Plots on the right show how the activity of each neuron changes due to these modes of bump kinetics. (C) Time series of μ and σ extracted from the fit. (D) Ongoing fluctuations in each bump parameter contribute a template matrix of $V_{\text{m}}$ covariances (color maps show covariances between cells with preferred orientation [PO] indicated on the axes of the “full” matrix, bottom right), obtained from (the outer product of) the differential patterns on the right of (B). Insets show $V_{\text{m}}$ covariance implied by each template for pairs of identically tuned cells (orange, PO difference $\simeq 0^{\circ }$) and orthogonally tuned cells (gray, PO difference $={90}^{\circ }$), as a function of stimulus orientation relative to the average PO of the two cells. The two templates sum up to a total covariance matrix (“bump kinetics”), which captures the key qualitative features of the full $V_{\text{m}}$ covariance matrix (“full”). The covariance matrix of the input noise (“input”) is also shown above for reference. The stimulus is at $0^{\circ }$ throughout.

Figure 6

Stimulus Tuning of Spike Count Correlations in the Ring SSN versus the Multi-attractor Ring Model (A) Spike count correlation matrix in the ring SSN during evoked activity (100% contrast). Color map shows correlations between cells with preferred orientation (PO) indicated on the axes, relative to stimulus orientation at $0^{\circ }$. Arrows indicate axes along which cell pairs are similarly (orange) or orthogonally tuned (gray). Spike count correlations along the diagonal show correlation for identically tuned cells, rather than for identical cells, and are thus less than one due to private spiking noise. (B) Average spike count correlations in the SSN, for pairs of similarly tuned cells (orange, PO difference less than ${45}^{\circ }$) and orthogonally tuned cells (gray, PO difference greater than ${45}^{\circ }$), as a function of stimulus orientation relative to the average PO of the two cells. (C and D) Same as (A) and (B), for the multi-attractor ring network. (E) Same as (B) and (D), for data from awake monkey V1 (Ecker et al., 2010). Data were symmetrized for negative and positive stimulus orientations. Shaded regions denote 95% CI. SSN simulations in this figure used the same parameters as in Figures 4 and 5.

Figure 7

Temporal Dynamics of Variability Modulation in the SSN versus Other Models (A) Time course of variability reduction and recovery in the ring SSN in response to a step input (shaded area, 500 ms duration) of increasing amplitude (dark to light). Variability is quantified by the population-averaged across-trial $V_{\text{m}}$ variance. (B) Same as (A), for chaotic network dynamics (Rajan et al., 2010). (C) Same as (A), for a continuous, multi-attractor network (Ponce-Alvarez et al., 2013). The stimulus is twice as long as in (A) and (B), so that variability suppression can be observed following the characteristic transient increase. (D) Timescale of variability suppression (time to reach half of the total suppression) as a function of the percentage of variance suppression in the three models, extracted from their corresponding variability trajectories (colors as in A–C). (E) Same as (D), for recovery timescales (time to recover half of the total suppression). In both (D) and (E), open yellow squares indicate V1 data from anesthetized cat (estimated from Figure S4 in Churchland et al., 2010); yellow circles show data from anesthetized monkey MT (Figure S4 in Churchland et al., 2010); dotted yellow circles show our analysis of the awake monkey V1 data of Ecker et al., 2010. Variability refers to the above-Poisson part of spike count variability (i.e., population-averaged Fano factor minus one), and time constants discard latencies in data. In the data of Ecker et al., 2010, the Fano factor dropped below one, effectively resulting in $>100\text{\%}$ variance suppression with our definition (right-pointing arrows). All results regarding the SSN and the multi-attractor model shown in this figure were obtained by using the same parameters as in previous figures (Figures 4, 5, and 6).