On the representation and learning of concepts : programs, types, and bayes

Abstract

This thesis develops computational models of cognition with a focus on concept representation and learning. We start with brief philosophical discourse accompanied by empirical findings and theories from developmental science. We review many formal foundations of computation as well as modern approaches to the problem of program induction - the learning of structure within those representations. We show our own research on program induction focused on its application for language bootstrapping. We then demonstrate our approach for augmenting a class of machine learning algorithms to enable domain-general learning by applying it to a program induction algorithm. Finally, we present our own computational account of concepts and cognition.