Friday, May 21, 2004

12 noon

Redwood Neuroscience Institute

 

Title:   “Learning Lateral Connections between Hidden Units”

 

Geoffrey Hinton

Computer Science Department

University of Toronto

 

Abstract:

I shall show how to approximate maximum likelihood learning in a generative model that consists of a hidden Markov Random Field with causal connections to the observed data. After choosing a hidden code to represent a datavector, the MRF runs for a few more iterations ignoring the likelihood term. This provides the statistics it needs to construct a complicated prior over code vectors. The prior uses lateral interactions, instead of independence, to squeeze the redundancy out of the code vectors, so the hidden units naturally form population codes.