Maximum-Entropy, Minimum-Description-Length Analysis of Associative Learning

October 6, 2014
Speaker: Charles Randy Gallistel, Professor, Rutgers Center for Cognitive Science, Rutgers University


The phenomena of associative learning are virtually universal in the animal kingdom. The wealth of data provide a proving ground for computational theories of cognition. We present an analysis of associative learning rooted in two information-theoretic principles of inference: maximum entropy and minimum description length. The minimum description length principle (MDL) provides a principled method of adjudicating the inescapable trade-off between model complexity and prediction error. It seeks the model that allows for the most compact encoding of the data. As a model becomes more complex, its ability to compress the data of experience increases, thus requiring fewer bits for data description, but only at the cost of requiring more bits to describe the model itself. By measuring both vices in a common currency (bits), MDL provides a principled method of selecting between alternative models. Given simple assumptions about how the inferred stochastic models translate into behavior, we demonstrate that a maximum-entropy-minimum-description-length model predicts a broad range of behavioral results from the literature on associative learning, including cue-competition, response-timing, extinction, inhibition, and partial-reinforcement results.

Speaker Bio

Randy GallistelRandy Gallistel is Professor of Psychology at Rutgers University, New Brunswick. He is a member of the Society of Experimental Psychologists, a Warren Medalist, and a member of the American Academy of Arts and Sciences and of the National Academy of Sciences. His research spans behavioral neuroscience and cognitive science, focusing on animal cognition, the nature of learning, the psychophysics of abstract quantities (space, time, number and probability), and the neurobiology of memory.




Hosted by Aurel A. Lazar

500 W. 120th St., Mudd Room 524, New York, NY 10027    212-854-5660               
©2014-2019 Columbia University