Representation of concurrent stimuli by population activity in visual cortex

Abstract

How do neuronal populations represent concurrent stimuli? We measured population responses in cat primary visual cortex (V1) using electrode arrays. Population responses to two superimposed gratings were weighted sums of the individual grating responses. The weights depended strongly on the relative contrasts of the gratings. When the contrasts were similar, the population performed an approximately equal summation. When the contrasts differed markedly, however, the population performed approximately a winner-take-all competition. Stimuli that were intermediate to these extremes elicited intermediate responses. This entire range of behaviors was explained by a single model of contrast normalization. Normalization captured both the spike responses and the local field potential responses; it even predicted visually evoked currents source-localized to V1 in human subjects. Normalization has profound effects on V1 population responses and is likely to shape the interpretation of these responses by higher cortical areas.

Figure 1

Tuning curves and population responses to single orientation stimuli in primary visual cortex. (A) Orientation tuning curves of all responsive sites (66 of 96), sorted according to preferred orientation. Each tuning curve is normalized by its mean across orientations. (B) Population response to a 45° stimulus: responses of all sites (dots) as a function of preferred orientation of each site. (C) The population response in B after binning sites with similar orientation preference (bin width: 15°). The curve is the best fitting circular Gaussian. Error bars indicate ± 1 SE of responses across sites in each bin. (D) Population response to a 0° stimulus for three contrasts: 12%, 50% and 100%. The abscissa indicates preferred orientation relative to stimulus orientation. Data for stimuli of multiple orientations (0°, 30°, and 60°) are combined to obtain each population response. Error bars indicate ± 1 SE of responses across sites in each bin. The curves fitting the data are circular Gaussians differing only in amplitude. (E) Amplitude of the population responses as a function of stimulus contrast. The curve is the best-fitting hyperbolic ratio function (c₅₀= 42.1%, n = 1.0). All fits are given by Equation (1). Experiment 84-12-16.

Figure 2

Relationship between single neuron (N = 75) tuning curves for orientation and contrast and response properties of the population. A Orientation tuning curves of single neurons (gray), normalized to the maximum response and centered on preferred orientation. Superimposed is the average orientation response profile of the population (black). B Contrast response functions. Same format as in A. C Semisaturation contrast (obtained by fitting Equation (2)) plotted as a function of tuning width (half-width at half-height) as obtained by fitting a circular Gaussian to each neuron’s response. Tuning width and semisaturation contrast are independently distributed across neurons in the population.

Figure 3

Population responses to plaids and predictions of the weighted-sum model. (A) Stimuli were gratings of two different orientations (first row and first column) and plaids obtained by summing the individual component gratings, for different combinations of component contrasts. Only a subset of stimuli used in the experiment is shown. (B) Population responses to these stimuli. Error bars indicate ± 1 SE of responses across sites in each bin. Model predictions for the single grating responses are given by the separable model of orientation and contrast (traces). (C–E) Predictions of the weighted-sum model for the population responses to plaids with (C) best-fitting weights, (D) equal weights, (E) winner-take-all weights. The data are re-plotted from (B) for comparison. Data in (B–E) are from experiment 83-10-15, plaid angle = 90 deg.

Figure 4

Predicted weights and fit quality of the weighted-sum model. (A) The complete stimulus set: gratings of different contrasts and their combinations (plaids). Population responses to plaids are fitted by the weighted sum of the responses to the component gratings (weighted-sum model, curves). Error bars indicate ± 1 SE of responses across sites in each bin. (B–G) Predicted weights (left) and average fit quality (right) for N = 9 experiments, for the weighted-sum model with best-fitting weights (B–C), equal weights (D–E), and winner-take-all weights (F–G) presented in the same format as (A). Plaid angle was 90 deg in 5 experiments, 45 deg in 2 experiments, and 30 deg in 2 experiments. Data in (A) are from experiment 82-6-3, plaid angle = 90 deg.

Figure 5

Predicted population responses, weights and fit quality of the normalization model. (A) Population responses are fitted by the normalization model with a single set of parameters. Error bars indicate ± 1 SE of responses across sites in each bin. (B) Weights predicted by the normalization model. (C) Average fit quality of the normalization model. Data in (A) are from experiment 82-6-3, plaid angle = 90 deg.

Figure 6

Competitive interactions between superimposed gratings as measured by large-scale subthreshold population activity. Curves are fits of the normalization model. Error bars indicate standard error across subjects. (A–B) LFPs obtained from area V1 in anesthetized cats in response to the test (A) and the mask stimulus (B). Response amplitudes were extracted at twice the stimulus frequency. Open symbols are conditions in which the test is presented alone, closed symbols indicate conditions in which a 25% contrast mask is added. (C–D) Source-imaged VEP signals obtained from area V1 in human observers using current source density modeling and fMRI retinotopic mapping. Format as in A–B. Responses in individual experiments were normalized to yield a value of 1 when the test stimulus was presented alone at 25% contrast.

Figure 7

Model of the effects of V1 normalization on a pattern-selective MT neuron. (A) Weights of a model MT neuron as a function of preferred direction of V1 neurons. (B) V1 population response (left) and response of the MT neuron (right, filled circle) to a single grating of 50% contrast drifting in a direction of 180 deg. Gratings of different directions of motion will elicit shifted versions of the V1 population response (not shown) which combine into the MT tuning curve (black line). (C) Responses to a plaid of equal component contrast (component directions are 120 and 240 deg, plaid motion direction is 180 deg). V1 population response (left), with normalization (black) and without normalization (gray). Response of the MT neuron to the plaid (right, filled circle). Plaids of different global motion directions will elicit shifted versions of the V1 population response (not shown), which are combined into the MT tuning for plaids (black trace: with V1 normalization; gray trace: without V1 normalization). The MT neuron responds most strongly when the global motion direction of the plaid is 180 deg. (D) Responses to the same plaid, but with unequal component contrast. Same format as in C. With normalization in V1, the MT neuron’s tuning for plaid direction shifts towards 240 deg, i.e. towards the plaid with directions of 180 deg (high contrast) and 300 deg (low contrast).