Richard
Zemel
Department
of
Computer
Science
University
of
Toronto
Neural Representations in a Dynamic Uncertain World
As animals interact with their environments, they must constantly update estimates about relevant states of the world. For example, a batter must rapidly re-estimate the velocity of a baseball as he decides whether and when to swing at a pitch. Probabilistic models provide a description of optimal updating based on prior probabilities, a dynamical model, and sensory evidence, and have proved to be consistent with the results of many diverse psychophysical studies. In this talk I will consider three basic questions. First, how can populations of neurons represent the uncertainty that underlies this probabilistic formulation? Second, how can neural spikes be used to represent adaptive computation? And finally, how can the population responses be learned from interaction with the environment? I will describe some recent progress in machine learning that suggests that predictions of behavior can be improved by incorporating particular constraints in a learning model.