Abstract

Short abstract: In this talk I’ll present our work on explaining motion perception as hierarchical causal inference. I’ll describe the intuitions behind the theory and show new psychophysics data task that quantitatively tests our theory. I’ll also describe our work in progress on using neural responses to test our theory, as well as Bayesian models of behavior in general.

Long abstract: If motion is always defined relative to a reference frame, what is the brain’s reference frame for the perception of a moving object? A century of psychophysical studies has provided us with seemingly conflicting evidence about motion perception in a variety of reference frames: from egocentric, to world-centric, to reference frames defined by other moving objects. We present a hierarchical Bayesian model which describes how observed retinal velocities give rise to perceived velocities. The hierarchically recurring generative model motif represents each perceived object’s motion in its natural reference frame which reflects the causal structure of the world. The degeneracy of object motion and reference frame motion is broken by a spike and slab prior reflecting the fact that most objects are exactly stationary in their natural reference frame. Data from three new psychophysical experiments quantitatively confirm key predictions of our model. Finally, I will present a stepwise method for generating neural predictions from our, and other, Bayesian models of the brain, and for comparing them against each other using neural data. Interestingly, a neural circuit implementing a generalized version of divisive normalization can generate the center-surround tuning curves predicted by causal inference.

Related manuscripts: Shivkumar, S., DeAngelis, G. C., & Haefner, R. M. (2023). Hierarchical motion perception as causal inference. https://doi.org/10.1101/2023.11.18.567582 Lengyel, G., Shivkumar, S., & Haefner, R. M. (2023). A General Method for Testing Bayesian Models using Neural Data. UniReps: The First Workshop on Unifying Representations in Neural Models. https://openreview.net/forum?id=oWJP0NhcY7