Poster Presentation

Poster E in Poster Session E: Thursday, August 6, 10:30 am – 12:15 pm, Kimmel Center, Shorin & Rosenthal Rooms

Reused representations and connectivity drive compositional generalization

Iman A. Wahle¹, Timothy J Buschman¹, Tatiana A Engel¹; ¹Princeton University

Presenter: Iman A. Wahle

Humans exhibit a remarkable capacity for few-shot generalization—acquiring new tasks from limited experience. This ability is thought to arise from compositional reuse of computations, with complex behaviors assembled from simpler components. Understanding the computational basis of this reuse could illuminate both how the brain achieves flexible learning and how to build artificial systems capable of the same. Compositional reuse could rely on the sharing of representations of task variables. Alternatively, the same connectivity could be reused across tasks. We do not yet know whether both forms of reuse occur in neural networks, and whether reuse of representations or connectivity predicts generalization. To systematically compare representation and connectivity reuse, we adapted an existing activity-based measure of reuse and introduce a novel connectivity-based measure that quantifies the extent to which circuitry is shared across tasks. To test these statistics, we applied them to minimal artificial reference networks with known degrees of compositional structure as standardized benchmarks. We found that the two measures diverge when task-irrelevant activity is present—suggesting that the choice of component is not merely definitional, but has measurable empirical consequences. We then correlated measures of reuse with the rate at which networks generalize to new tasks. We found both representation and connectivity reuse each significantly predict generalization speed. Together, these results demonstrate that while reuse can occur independently at the levels of activity and connectivity, both may support generalization and could serve as inductive biases for flexible, few-shot learning in artificial systems.

Topic Area: Methods, Tools, Theory & Neural Coding