2:00pm, Tuesday, 17 July 2007:

<center><b>
Encoding Probability Distributions in the Visual Cortex
</b>

<p>
<a HREF="http://www.cs.cmu.edu/~yan/">
Yan Karklin</a>
<br>
Carnegie Mellon University

</center>



A fundamental function of the visual system is to encode the building
blocks of natural scenes -- edges, textures, and shapes -- that subserve
tasks like object recognition and scene understanding.  Essential to this
process is the formation of abstract representations that are stable
despite the wide natural variability in the appearance of individual image
elements.  We argue that, in order to achieve this, neurons in the visual
cortex must encode entire patterns of variation in visual stimuli, rather
than signalling the presence or conjunction of a small number of image
features.  We have developed a statistical model that implements this
idea; in the model, the joint activity of neurons encodes a probability
distribution over the inputs and describes complex patterns of variation
associated with them.  Trained on an ensemble of natural images, the model
learns a compact set of functions that act as dictionary elements for
distributions typically encountered in natural scenes.  Units in the model
exhibit striking similarities to properties observed in V1 and V2 neurons,
suggesting new ways to analyze neural codes in the visual cortex.