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

Poster C in Poster Session C: Wednesday, August 5, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Learning Long-Range Temporal Dependencies Through Sleep-Inspired Acceleration

Nicholas Soures¹, Miranda Gonzales¹, Abinidhi Geethaikrishnan¹, Itamar Lerner¹, Dhireesha Kudithipudi¹; ¹University of Texas at San Antonio

Presenter: Nicholas Soures

Learning long-range dependencies between events spread out in time remains difficult for systems constrained by finite memory, posing a central challenge for continual learning systems. In biological neural networks, the temporal scaffolding hypothesis asserts that memory replay during slow-wave sleep functionally reduces the temporal distance between associated events, facilitating novel associations to emerge in subsequent processing. Inspired by this principle, we investigate whether accelerating the propagation of sequential data during learning can improve learning across long horizons. Our results suggest that learning from temporally accelerated sequences provides a complementary mechanism to extend the effective depth of memory beyond the inherent context window and the potential to integrate with existing learning rules.

Topic Area: Memory, Learning & Knowledge Structures