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

Poster F in Poster Session F: Thursday, August 6, 1:45 – 3:30 pm, Kimmel Center, Shorin & Rosenthal Rooms

Continual learning and forgetting as drivers of representational drift

Chattarin Poungtubtim¹, Timothy Brady¹, John T Serences¹; ¹University of California, San Diego

Presenter: Chattarin Poungtubtim

Recent work demonstrates that stimulus- and task-related population activity reorganizes over days to weeks even when behavioral performance appears stable representational drift. While theoretical work suggests that random diffusion (i.e. noise) may be sufficient to explain drift, recent studies also suggest that task representations become sparser and more separable over time. This latter observation is consistent with the hypothesis that representational drift is partially driven by ongoing learning rather than purely stochastic processes. Here, we use continuous time recurrent neural networks (ctRNNs) to demonstrate that over-training - even after reaching criterion accuracy - leads to changes in activity patterns similar to representational drift. We also find that task representations in the network become sparser and more separable over time, and that these changes result in greater tolerance to perturbations such as increases in internal and external noise. Our results suggest that drift can reflect continual learning and optimization as opposed to just random diffusion and is characterized by increases in the sparsity, efficiency, and robustness of neural codes.

Topic Area: Methods, Tools, Theory & Neural Coding