Self-Organizing Neural Networks And Topographic Map Formation In The Cerebral Cortex

Friday, March 28, 1997
11 a.m. - 12 noon
Fuller Labs 320

Dr. Sergio A. Alvarez
Center for Nonlinear Analysis
Carnegie Mellon University

Neural networks (NN) are canonical examples of connectionist AI models that exhibit several interesting and useful properties such as self- organizing behavior, even when equipped only with extremely simple unsupervised learning rules, and robustness with respect to damage involving significant portions of the system. From a philosophical viewpoint, NN are appealing because they constitute simple computational models of the operation of the brain. In this talk I will focus on this latter aspect.

I will begin by reviewing some of the basic concepts of NN. I will then work with an NN model of the cerebral cortex having two distinct "hemispheres" connected by a simulated "corpus callosum," and provided with input from a simulated "sensory organ." I will describe topographic map formation in such a system. This is a special type of self-organizing behavior in which a system with initially random interconnection weights and subjected to randomly chosen sensory stimuli is able to learn the geometry of its own sensory organ. I will show that the resulting "maps" exhibit a variety of organization patterns across the two hemispheres, and in particular provide a computational demonstration of the well-documented neurological phenomenon of cerebral lateralization.

This is joint work with J. Reggia and S. Levitan, Department of Computer Science, University of Maryland at College Park.

Maintained by webmaster@wpi.edu
Last modified: Sep 27, 2006, 16:05 EDT