Predictive Coding for Unsupervised Feature Learning

Summary

Date: 
December 14, 2016 - 1:00pm - 2:00pm
Location: 
Northwest 243
About the Speaker
Name: 
Bill Lotter
Speaker Affiliation: 
Harvard, MCB

Abstract:
"Predictive coding" is an influential theoretical framework in neuroscience that posits that populations of cortical neurons actively attempt to predict incoming sensory information, and that only deviations from those predictions are sent on for subsequent processing in a cortical hierarchy. While predictive coding is often framed as a scheme for efficient coding, here, we explore prediction as a rule for supervised learning. In particular, we bring predictive coding together with modern deep learning techniques to build a network that can predict future frames of a video, motivated by the idea that the ability to predict future frames requires a robust internal representation of object structure and an understanding of the rules by which the world changes. We demonstrate that these networks can learn to predict the motion of synthetic stimuli and, in doing so, learn a representation of the underlying latent parameters, which is useful for object recognition. We show that these networks can also scale to complex natural videos, where they accurately predict future frames in car-mounted camera videos, successfully estimating egocentric motion and the motion of other objects in the scene. These results support the use of prediction as a powerful rule for unsupervised learning.