Bayesian Models of Conceptual Development: Learning as Building Models of the World
Tomer David Ullman, Joshua B. Tenenbaum · 2020
A Bayesian framework helps to address, in computational terms, what knowledgechildren start with and how they construct and adapt models of the worldduring childhood. Within this framework, inference over hierarchies of probabilisticgenerative programs in particular offers a normative and descriptiveaccount of children's model-building. We consider two classic settings in whichcognitive development has been framed as model-building: (i) Core knowledgein infancy, and (ii) The child as scientist. We interpret learning in both of thesesettings as resource-constrained, hierarchical Bayesian program induction withdifferent primitives and constraints. We examine what mechanisms childrencould use to meet the algorithmic challenges of navigating large spaces of potentialmodels, in particular the proposal of \the child as hacker" and how itmight be realized drawing on recent computational advances. We also discussprospects for a unifying account of model building across scientific theories andintuitive theories, and in biological and cultural evolution more generally.