Seminar on Computational Learning and Adaptation

and

NOBOTS

 
A Bayesian Framework for Concept Learning

Josh Tenenbaum
Department of Psychology
Stanford University
jbt@psych.stanford.edu

I will describe a Bayesian framework aimed at understanding human concept learning in computational terms and bringing machine concept learning systems closer to the potential of human learners. I will present theoretical analysis and data from several experiments with human subjects that address three specific questions. How are people often able to generalize a concept from only a small number of positive examples?  How does people's prior knowledge interact with the examples they observe to guide the generalizations that they make?  Why does generalization appear in some cases to be based on abstract rules and in other cases to be based on similarity to exemplars?  I will also discuss some implications for machine concept learning systems, with a focus on the problem of example-based database retrieval.

Return to the seminar schedule