Coding for Uncertainty

Peter Dayan
University College London
Gatsby Computational Neuroscience Unit
United Kingdom

Abstract

Perceptual inference fundamentally involves uncertainty, arising from noise in sensation and the ill-posed nature of most perceptual problems. Accurate perception requires this uncertainty to be represented, manipulated and learned about correctly, even as it changes dynamically. Uncertainty also needs to be distinguished from potential confounding factors such as multiplicity or transparency. In this talk, I will discuss our investigations into how populations of neurons can offer implicit and explicit representations of various forms of uncertainty. This is joint work with Rich Zemel, Alex Pouget, Quentin Huys and Rama Natarajan.

Return to the Theoretical Neuroscience Symposium Main Page