2015-10-07 at 11:00
Conference room, UCL R+0, 13 rue Moreau, 75012 Paris
How the statistics of the sensory environment shape cortical visual processing
Several decades of work have suggested that Barlow's principle of efficient coding is a powerful framework for understanding retinal design principles. Whether a similar notion extends to cortical visual processing is less clear, as there is no "bottleneck" comparable to the optic nerve, and much redundancy has already been removed by the retina. Here, we present convergent psychophysical and physiological evidence that regularities of high-order image statistics are indeed exploited by central visual processing, and at a surprising level of detail. We recently showed that high-order correlations in certain specific spatial configurations are informative, while high-order correlations in other spatial configurations are not: they can be accurately guessed from lower-order ones. We then construct artificial images (visual textures) composed either of informative or uninformative correlations. We find that informative high-order correlations are visually salient, while the uninformative correlations are nearly imperceptible. Physiological studies in macaque visual cortex identify the locus of the underlying computations. First, neuronal responses in macaque V1 and V2 mirror the psychophysical findings, in that many neurons respond differentially to the informative statistics, while few respond to the uninformative ones. Moreover, the differential responses largely arise in the supragranular layers, indicating that the computations are the result of intracortical processing. We then consider low- and high-order local image statistics together, and apply a dimension-reduction (binarization) to cast them into a 10-dimensional space. We determine the perceptual isodiscrimination surfaces within this space. These are well-approximated by ellipsoids, and the principal axes of the ellipsoids correspond to the distribution of the local statistics in natural images. Interestingly, this correspondence differs in specific ways from the predictions of a model that implements efficient coding in an unrestricted manner. I suggest that these deviations provide insights into the computational mechanisms that underlie the representation of image statistics.