Abstract:

How do we learn so much from so little? A chef tastes one spice and knows which dishes it will complement; a scientist proposes new theories and new experiments from a handful of observations. I argue that structured mental representations play a key role: instead of slowly accumulating statistics, the mind builds compact hypotheses and recombines them strategically. In this talk, I will show how Bayesian program induction can model these structured representations across diverse learning tasks—generalizing causal relations from single observations, building compositional concepts through curricula, and planning exploration in gridworld games. Our models accurately capture human behavior and systematically outperform unstructured alternatives. This work suggests that human-like AI may require not bigger datasets, but smarter representations—the kind that let us learn the rules of the game, not just how to play it.