THIS PAGE INCORPORATES CASCADING STYLE SHEET! TO VIEW PAGE CORRECTLY PLEASE UPDATE TO A CSS COMPLIANT BROWSER
 |
The Visual Cortext as a Hierarchical Predictor |
pdf 204 KB
ps.gz 225 KB
ps 817 KB
@TechReport{Rao:TR96,
author = {Rajesh P.N. Rao and Dana Ballard},
title = {The Visual Cortext as a Hierarchical Predictor},
institution = {CS Dept. University of Rochester},
year = {1996},
number = {96.4},
address = {Rochester NY},
month = {September},
abstract = { A fundamental feature of the visual cortex is the reciprocity of connections between its
many distinct areas. Using the principle of Kalman filtering from classical control theory,
we describe how this reciprocity may allow the cortex to function as a hierarchical predictor.
Feedback connections in this model carry predictions of lower level neural activities, while
feedforward connections convey to the higher level the differences between the predictions
and the actual lower level activities. These differences allow the higher level to maintain
optimal estimates of current visual recognition state and on a longer time scale, enable it to
learn a dynamic internal model of the visual environment. A simulated Kalman filter network
embodying these principles produced response properties that correlated closely with those
of neurons in the primary visual cortex. }
}