Bayesian Models of Conceptual Development: Learning as Building Models of the World

Abstract

A Bayesian framework helps address, in computational terms, what knowledge children start with and how they construct and adapt models of the world during childhood. Within this framework, inference over hierarchies of probabilistic generative programs in particular offers a normative and descriptive account of children's model building. We consider two classic settings in which cognitive development has been framed as model building: ( a) core knowledge in infancy and ( b) the child as scientist. We interpret learning in both of these settings as resource-constrained, hierarchical Bayesian program induction with different primitives and constraints. We examine what mechanisms children could use to meet the algorithmic challenges of navigating large spaces of potential models, in particular the proposal of the child as hacker and how it might be realized by drawing on recent computational advances. We also discuss prospects for a unifying account of model building across scientific theories and intuitive theories, and in biological and cultural evolution more generally.