Of Flies and Men: How Natural Stimulus Correlations Influence Visual Motion Estimation.

The estimation of visual motion is a paradigmatic neural computation, and multiple models have been advanced to explain behavioral or neural responses to motion signals. A broad class of models, epitomized by the Reichardt correlator model, proposes that animals estimate motion by computing a cross-correlation of light intensities from two neighboring points in space. In specific contexts, these models adequately describe experimental data across diverse organisms, such as flies and humans, but these models do not generalize to arbitrary visual stimuli. Here, we describe a theoretical formalism that we developed by treating motion estimation as a problem of Bayesian inference. While Reichardt models emerge as one component of the generalized motion estimation strategy, more complex correlation types are required for accurate motion estimation. In particular, correlations of both even and odd orders are predictive of motion within the natural world. Interestingly, classical experiments using standard laboratory stimuli cannot reveal whether the subject uses odd-order correlators for motion estimation. Our theory demonstrates how correlation-based motion estimation is related to stimulus statistics and provides multiple experimentally testable predictions. One key prediction of the theory is that complex spatiotemporal stimulus correlations can induce motion percepts. Testing this experimentally is complicated by pairwise correlations that generate a motion signal from the Reichardt correlator. Yet, novel stimuli that overcome this difficulty indicate that third and fourth-order correlations induce motion percepts in humans. We are now using these stimuli to test the prevailing view that the Reichardt correlator is the only motion estimator in Drosophila.