I will introduce a stochastic lambda calculus, and its implementation as the probabilistic programming language Church. Certain expressions in Church can be seen as generalized probabilistic grammars. I will then argue that this is a useful formalism for understanding human concepts, and hence concept learning can be seen a process of probabilistic program induction. I will illustrate this with a model of the acquisition of natural number concepts, and an experiment investigating the ability of participants to learn recursive and constituent structure from tree-like stimuli. Time permitting, I will describe a second approach to learning abstractions, called Fragment Grammar, based on stochastic reuse of computation.

Concept Learning as Program Induction

Noah Goodman

MIT