Reproduce “Leakage and Second-Order Dynamics Improve Hippocampal RNN Replay” on OpenResearch

About this paper

Josue Casco-Rodriguez, Nanda H. Krishna, Richard G. Baraniuk

The authors show how a noisy recurrent neural network can replay path-integrating activity after self-motion cues stop. They add a momentum-like state to adaptation-based replay, allowing earlier changes to persist, and analyze leakage as a time-varying score representation. In simulations, underdamping reaches T-maze endpoints, defined as within 10%, faster than awake paths while average exploration is maintained, not established biologically.

Key insights

• Optimal path-integration in noisy environments implies a time-variant, linear score function, providing a theoretical rationale for linear leakage in recurrent neural networks.
• Neural adaptation introduces second-order dynamics that aid exploration in replay but inherently reduces replay speed and can decrease fidelity to the original experience.
• Explicitly incorporating a momentum term to achieve underdamped Langevin dynamics effectively enables temporal compression in replay, accelerating sequences without compromising exploration.

Summary by alphaXiv.