Emergence of a contrast-invariant representation of naturalistic texture in macaque visual cortex.

UNLABELLED

Sensory stimuli vary across a variety of dimensions, like contrast, orientation, or texture. The brain must rely on population representations to disentangle changes in one dimension from changes in another. To understand how the visual system might extract separable stimulus representations, we recorded multiunit neuronal responses to texture images varying along two dimensions: contrast, a property represented as early as the retina, and naturalistic statistical structure, a property that modulates neuronal responses in V2 and V4, but not in V1. We measured how sites in these 3 cortical areas responded to variation in both dimensions. Contrast modulated responses in all areas. In V2 and V4, the presence of naturalistic structure both modulated responses and increased contrast sensitivity. Tuning for naturalistic structure was strongest in V4; tuning in both dimensions was most heterogeneous in V4. We measured how well populations in each area could support the linear readout of both dimensions. Populations in V2 and V4 could support the linear readout of naturalistic structure, but only in V4 did we find evidence for a robust representation that was contrast-invariant.

SIGNIFICANCE STATEMENT

Single neurons in visual cortex respond selectively to multiple stimulus dimensions, so signals from single neurons cannot distinguish changes in one dimension from changes in another. We measured responses from simultaneously recorded neural populations in three hierarchically linked visual areas - V1, V2, and V4 - using texture stimuli that varied in two dimensions, contrast and naturalistic image structure. We used linear decoding methods to extract information about each dimension. In all three areas, contrast could be decoded independently of image structure. Only in V4, however, could image structure be decoded independently of contrast. The reason is that selectivity for texture and contrast in V4 was much more diverse than in V1 or V2. This heterogeneity allows V4 to faithfully represent naturalistic image structure independent of contrast.