Title: Multi-layer DCA, Data Manifolds, Statistical Physics and Visual Processing in the Brain

 

Tony Bell

Redwood Neuroscience Institute

 

Abstract:

I will discuss high-dimensional probability density estimation from the point of view of Dependent Component Analysis (DCA). First I will show the general structure of this algorithm. Then I will show results of 'Subspace ICA' and 'Topographic ICA' (two DCA algorithms)on natural images, showing, as Hyvarinen & Hoyer did, how the results relate to the computation of area V1 of cortex. Next, I will show how to turn Topographic DCA into a multi-layer algorithm, and hopefully present results. Finally, I will address the (ever present) relations with statistical physics and manifold interpolation and point to some intriguing research directions.