Abstract
Perceptual spaces play a key role in intermediate stages of visual processing, as they constitute representations that support discrimination, categorization, working memory, and other judgments. The classical perceptual space of color has three dimensions; others, such as faces and textures, have very high dimension. Representing such spaces within biological constraints is challenging. To probe how visual cortex does this, we studied responses of macaque single neurons to a well-characterized 10-dimensional domain of visual textures. This model space consists of black-and-white colorings of a checkerboard, and its spatial structure is defined by multipoint correlations within 2x2 blocks (Victor and Conte 2012). The 10 parameters of the space are: a statistic that specifies the balance of white vs. black checks, four pairwise statistics that specify nearest-neighbor correlations, four three-point statistics that specify the prevalence of triangular regions, and a four-point statistic. This space captures the informative local image statistics of natural images (Hermundstad, 2014). Human perceptual sensitivities within the space are well-characterized, and suggest opponent mechanisms that combine via quadratic summation. We recorded from 214 neurons via tetrodes in V1 and V2 of 6 anesthetized macaques. Neurons responsive to one- and two-point statistics were common (>50%) in V1 and V2, with the previously-noted bias towards darks in both regions (Yeh et al., 2009). Responses to three- and four-point statistics were rare (< 10%) in V1, but more common in V2 (~20%). Consistent with an opponent representation, neurons that increased their firing in response to variation in one direction along a coordinate axis tended to decrease their firing in response to variation in the opposite direction. However, typical neurons responded to variations along several coordinate axes, with no clear pattern of "cardinal" directions. Thus, our data suggest a multidimensional opponent representation for local image statistics in the population of V1 and V2 neurons.
Meeting abstract presented at VSS 2015