Learning Sparse Representations for Vision

Tomaso Poggio
Uncas and Helen Whitaker Professor
Brain Sciences Department and A.I. Lab
M.I.T.

 

Abstract

I will describe some recent results on the problem of function approximation and sparse representations that connect regularization theory, Support Vector Machine Regression (Vapnick), Basis Pursuit Denoising (Chen, Donoho, Sanders) and PCA techniques. I will motivate the appeal of learning sparse representations from an overcomplete dictionary of basis functions in terms of recent results in two different fields: neuroscience and computer vision. In particular, physiological data from IT cortex suggest that individual neurons encode a large vocabulary of elementary shapes before converging on cells tuned to specific views of specific 3D objects. In the area of computer vision we have developed a trainable object detection architecture that succeeds in learning a sparse representation from an overcomplete set of Haar wavelets to perform difficult object detection tasks.



Luis Gravano
gravano@cs.columbia.edu