Conference on Cognitive Computational Neuroscience

Humans have a remarkable ability to compositionally generalize, to recompose familiar structural components in new ways to solve new tasks. For instance, navigation tasks can be decomposed into two components: knowledge of the goal location, and knowledge of vehicle operations (e.g. pedaling a bike or driving a car) to reach the goal. Compositional generalization requires components to be independently transferable -- different vehicles can reach the same goal, and the same vehicle can reach different goals. Yet transferring entire task structures can be more adaptive, especially when these recur across multiple tasks (e.g., given a suggestion to get ice cream, one might prefer to bike, even in new towns). A satisfactory account of how an agent can transfer individual components but also entire structures is lacking. Here, we propose an agent that learns and transfers individual components as well as entire structures by inferring both through a non-parametric Bayesian model of the task. It maintains factorized representations of components but also represents different possible covariances between them. Our agent generalizes better across a variety of navigation tasks covering a range of component covariances, including hierarchical tasks with goal/subgoal structures. Finally, we discuss how cortico-striatal gating circuits could implement our algorithm.