Brain and Mind Institute Seminar
08.06.09 Thursday, 15h00, AAB032
Neural evidence for statistically optimal inference and learning in
primary visual cortex
Abstract:
How do we infer from sensation the state of the external world? Human
and animal subjects are able to take into account noise and
uncertainty in behavioral task and perform statistically optimal
inference and learning. Moreover, statistical models of natural images
have been shown to reproduce many features of receptive field
organization in primary visual cortex. However, there has been so far
no evidence of optimal inference and learning at the neural level. In
this talk, I will derive a general consequence of the statistical
framework, predicting that the distribution of neural spontaneous
activity and that of activity evoked by natural stimuli must become
more and more similar with visual experience, and be identical in the
ideal case, under the assumption that neural activity represents
samples from an internal, probabilistic model of the environment. I
will present data from multielectrode recording in awake ferrets at
various stage of post-natal development that supports this prediction.
The increasing similarity between the two distributions is found to be
due to an increasing match between the spatial and temporal
correlational structure of the activity patterns, and is specific to
activity evoked by natural stimuli, and not by noise or grating
stimuli. These results provide support for the statistical framework
at the neural level, and suggest a novel interpretation for neural
variability and spontaneous activity.
back |