Chris Rozell | Dimensionality Reduction as a Model of Efficient Coding in the Visual Pathway
The engineering and applied math communities often exploit the fact that natural stimuli have significant structure that lends itself well to dimensionality reduction. The efficient coding hypothesis for sensory neural coding postulates that stages of neural processing should sequentially make the representations more efficient by removing stimulus redundancies, and this is often expressed in the language of information theory. In this talk I will present our work exploring efficient coding models of vision based on dimensionality reduction, including sparsity, low-rank matrix factorizations and random projections. I will show that such approaches are able to account for many observed properties in visual cortex, including classical receptive fields, response properties based on nonclassical or nonlinear receptive fields, and properties of the inhibitory interneurons. NEUROSCIENCE WORKSHOP: Dimensionality Reduction Friday, October 30, 2015 Saturday, October 31, 2015