Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

A mathematical theory of hippocampal repulsion

Samuel Lippl¹, Christopher S. Iyer¹, Kim Stachenfeld¹, Anna C Schapiro²; ¹Columbia University, ²University of Pennsylvania

Presenter: Samuel Lippl

Flexible behavior requires both generalization across related stimuli and differentiation between them, processes that both implicate the hippocampus. While there are various theories of how the hippocampus may shape representations to accomplish these goals, there is an enduring unsolved mystery in this field: hippocampal representations sometimes become *more dissimilar* for stimuli (A1 and A2) with a shared associate (B) than stimuli with no shared associate—a counterintuitive phenomenon known as "repulsion." Repulsion has been suggested to be inconsistent with classical associative learning rules, which are commonly understood to lead to integration of related stimuli. Here we show that, surprisingly, a simple autoencoder exhibits repulsion: if the network uses B to predict both A1 and A2 at different times, the representations of A1 and A2 naturally come to inhibit each other. In contrast, a network that predicts B without reconstructing A1 and A2 yields integration. We analytically characterize the transition between integration and repulsion in a family of linear neural networks trained on a mixture of predictive and autoencoding objectives. We then show that autoencoders replicate key experimental findings of repulsion, establishing a novel, parsimonious account of representational shaping in the hippocampus.

Topic Area: Memory, Learning & Knowledge Structures