Probing the early visual system with naturalistic, synthetic images

Abstract

Event Abstract Back to Event Probing the early visual system with naturalistic, synthetic images The physiology of early visual cortex has traditionally been studied with simple stimuli such as bars and gratings. While this approach has revealed much of what we know, it is also likely to have provided an incomplete view of early visual processing in natural vision. There has thus been a growing interest in using natural images, namely photographic depictions of natural environments, to study cortical processing. A frequent criticism of this approach, however, is that natural images are so complex that it is difficult to understand precisely which of their features drive the observed physiological effects. Our aim is to develop a new class of stimuli that captures much of the statistical richness of natural images while still allowing full parametric manipulation. We have combined the development of this method with its application to neurophysiological experiments in primary visual cortex and with computational modeling of the experimental data. Our approach is based on characterizing the statistics of natural scenes in the context of an algorithm that generates synthetic images [1]. The algorithm is based on a steerable pyramid of oriented filters, which has a clear biological correlate in the receptive field properties of neurons in primary visual cortex (V1). The resulting parameter space is quite high dimensional. Here we focused on the joint statistics of the magnitude of filter activations, and defined two perceptually relevant parameters: H, the entropy of filter activity across orientations, which measures how well defined the average orientation of the image is; and L, the spatial extent of correlation among parallel filters, capturing the extent of edges and lines in the image. We analyzed an ensemble of natural scenes to find the parameter ranges typically encountered in such scenes and then generated artificial images by choosing values from each of the distributions. Interestingly, we found that natural images tend to cluster in a corner of the space defined by our parameters, that has little overlap with sinusoidal gratings. We synthesized images, varying H and L independently to obtain a parametric set of synthetic, naturalistic images. To test the applicability of our approach, we recorded the responses of single units in V1 of anesthetized macaque monkeys to synthetic stimuli presented either to the classical receptive field or extending into the non-classical receptive field (nCRF). The nCRF experiments showed a rich variety of nonlinear effects, including: a) clear dependency of nCRF surround suppression on the parameter H, and, to a less extent, on L ; and b) reduced suppression, and in some cases facilitation, when the center and the surround of the cells’ receptive field were stimulated with different statistics. Gain control models of cortical processing have been effective in explaining some V1 nonlinearities in response to grating stimuli; we are testing the ability of Gaussian Scale Mixtures, a class of generative model of natural image statistics related to divisive gain control, to account for our data.