Interplay between sensory statistics and neural circuit constraints in visual motion estimation


March 26, 2014 - 1:00pm
NW 243
About the Speaker
James Fitzgerald (Swartz Fellow)

Many animals use visual signals to estimate motion. Canonical models suppose that animals estimate motion by computing pair correlations between spatiotemporally separated visual signals, but recent experiments indicate that humans and flies also perceive motion from certain higher-order correlations that signify motion in natural environments. Here we study how nonlinear visual processing enables the computation of higher-order stimulus correlations. In particular, we frame visual motion estimation as a problem of constrained statistical inference and consider several candidate constraints. Interestingly, our results reveal two ways that efficient visual coding could improve motion estimation. First, contrast equalization transforms visual input streams in a manner that makes their cross-correlation more predictive of motion. Second, segregating motion signals into ON/OFF channels enhances estimation accuracy. Overall, our results provide further theoretical rationale for higher-order correlations in motion estimation and demonstrate how brains might extract such signals.