Thursday, January 15, 2004

12 noon

Redwood Neuroscience Institute

Title: Confabulation: The Basic Building Block of Thinking

Robert Hecht-Nielsen

Computational Neurobiology Program

Institute for Neural Computation

ECE Department

University of California, San Diego

Abstract:

This talk will discuss an idealized symbolic predictive modeling process called confabulation, which is hypothesized by the speaker to be the elemental information processing operation of thought (Hecht-Nielsen, R., A Theory of Cerebral Cortex (2003) Technical Report #0301, UCSD Institute for Neural Computation (ftp from inc2.ucsd.edu), Hecht-Nielsen, R. and McKenna, T. [Eds] (2003) Computational Models for Neuroscience, Springer-Verlag). Abbreviated confabulation maps a set of individual symbols, called assumed facts (each belonging to a different lexicon of symbols) to a set of symbols (called an expectation) belonging to yet another lexicon. Confabulation is akin to pattern classification, with the differences being that the input is a set of symbols from disjoint lexicons (not a point in Euclidean space) and the output is not a single ‘class symbol,’ but a set of such symbols. The expectation represents the set of all reasonable predictions, given the assumed facts. Just as with pattern classification, there is a theoretically ideal predictor: one based on omniscient knowledge (knowledge of all probabilities involving all permutations of all combinations of symbols). Confabulation is based upon a very limited type of knowledge (a collection of pairwise symbol conditional probabilities, which is assumed to be exhaustive). The Expectation Theorem (which will be presented and proven) shows that confabulation produces an expectation which contains the ideal expectation as a subset. Full confabulation, an add-on operation to abbreviated confabulation, produces a positive estimated quality value for each expectation element. The elements with the highest levels of estimated quality are termed answers. Computer experiments are presented which illustrate the efficacy of confabulation. A sketch will be presented of how the knowledge required by confabulation is acquired and stored in the cerebral cortex. The thalamocortical neuronal circuit that implements confabulation will also be discussed.

Speaker

Robert Hecht-Nielsen is with the Computational Neurobiology Program, the Institute for Neural Computation, and the ECE Department of the University of California, San Diego. An IEEE Fellow, he has received the IEEE Neural Networks Pioneer Award and the ECE Department’s Graduate Teaching Award.