Abstract
Image encoding in visual cortex is thought to be adapted to natural scene statistics. Spatial attention may selectively enhance the acquisition of these statistics. The effect of such attention-gated learning may be especially pronounced in peripheral vision, as the shift of attention to a peripheral visual location is tightly coupled to the initiation of a saccade. We hypothesize that this attention-saccade link biases peripheral visual input, introducing an eye-movement-related confound to the veridical image statistics (Nandy & Tjan, 2012). This, in turn, is predicted to shape learned spatiotemporal receptive fields and lateral interactions in ways that predict deficits of peripheral form vision such as crowding. To model the effect of saccade-confounded statistics on image encoding in V1, we simulated saccades across natural image data and mapped the resulting videos to a model cortical surface with eccentricity-dependent magnification and resolution. We then learned a space-time sparse coding dictionary on samples from a circular patch of model cortex – representing a target location and a laterally-interacting surround, all under an attentional spotlight centered on the target – in a temporal window adjusted to the duration of a transient spatial attention episode (Sperling & Weichselgartner, 1995). This peripheral spacetime dictionary is spatially similar to foveal dictionaries, but with a preference for movement toward the fovea. For a given basis trained on peripheral inputs, only a subinterval in time contains a structured pattern. This subinterval can appear at any time in the basis. The unique features of this dictionary – anisotropic movement preference and a narrow temporal window with variable onset delay – point toward a role for peripheral vision that may differ more substantially from foveal vision than previously thought.
Meeting abstract presented at VSS 2016