Title: Image Statistics,
Hyper-sparse Codes and the Geometry of
Non-linearities in Visual Neurons.
David Field
Department of Psychology
Abstract:
Recent work has demonstrated
that many aspects of visual coding can be accounted for in terms of the sparse
structure of natural scenes. However, the current approaches still account for
only a fraction of the dependent structure in images and account for only a
small portion of the known response properties of visual neurons. In this talk,
we consider some of the dependent structure of images in terms of sparse and
hyper-sparse codes (local 'winner take all' codes). It is proposed that by
considering the response geometry of these codes, we can gain significant
insights into the non-linear response properties of visual neurons.