Richard Lange
Department of Brain and Cognitive Science
Department of Computer Science
University of Rochester

Signatures of Approximate Bayesian Inference in Early Visual Processing

Perception is often cast as an inference process in which the objects, relations, and events we perceive are latent variables whose values must be inferred from their effects on the senses. At the core of this theory lie two unanswered questions: what is the nature of the brain’s internal model, and how are probability distributions of variables in this model represented? Previous research often takes a “forward engineering” approach to these questions, assuming a model family and inference algorithm and mapping the inference dynamics, theoretically or empirically, onto neural circuits. My research has instead been driven by a “reverse engineering” approach, making empirical predictions for Bayesian inference in perception that are not committed to a particular model family or inference algorithm. I will show how a highly general descriptive model of probabilistic representation is sufficient to make testable predictions for a general class of “posterior coding” representations, specifically for neural and behavioral patterns in the common two-alternative forced-choice discrimination paradigm, and I will present new data from both domains.

A pizza lunch will be served.