Title: "Using ICA to model
neural responses in V1"
Patrik Hoyer
Center for Neural Science,
Abstract:
What is the computational
purpose of the oriented, bandpass receptive fields of
simple cells in primary visual cortex (V1)? One suggestion is that these cells
decompose their natural input into approximately independent components. There
are two main theoretical frameworks in which to interpret this behavior. The
first is the ‘efficient coding hypothesis’, which postulates that neurons
communicate information in a maximally efficient manner. The second is the idea
that cortical circuits are inferring the hidden variables of a generative
model. I will start by briefly reviewing these two frameworks, and then move on
to describe how cortical ‘noise’ might be understood in the context of the
latter one, as a natural consequence of the inference process.