| ||
Beyond independence and sparseness in models of natural image statistics | ||
Aapo Hyvarinen | ||
Helsinki University of Technology | ||
A fundamental model of natural image statistics is independent component analysis which is essentially similar to linear sparse coding. In this talk, I will discuss some extensions and alternatives to these models. First, modeling the dependencies of the simple cell outputs leads to models of complex cells and topography. Second, I consider natural video data, and replace the sparseness criterion by temporal correlations. This leads to similar properties than sparseness: advantages of the two approaches are considered. | ||