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

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

Discovering Online and Offline Processing via Meta-Learning

Zhenglong Zhou¹, Marcelo G Mattar¹; ¹New York University

Presenter: Zhenglong Zhou

Biological memory systems employ distinct processes during wake and sleep, but how encoding and offline processing should be jointly organized remains an open question. We introduce a meta-learning framework in which a recurrent neural network jointly optimizes its weights and separate online and offline plasticity rules, so that meta-learning shapes both the network's dynamics and how it updates its own synapses. The model learns to rapidly adapt to novel task instances, and the meta-learned offline processing proves beneficial: removing it degrades performance, and substituting the learned offline plasticity with the learned online plasticity also impairs performance, indicating functionally distinct rules. The model also discovers reward-modulated plasticity without any explicit reward-dependent mechanism. Furthermore, the learned offline dynamics differ across tasks with distinct demands: tasks requiring sequential action sequences to reach reward elicit offline activity with positive similarity to online representations, whereas simpler tasks do not. These emergent properties suggest that meta-learning can reveal organizational principles of biological memory systems.

Topic Area: Memory, Learning & Knowledge Structures