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

Experience shapes strategy via embedded network dynamics

John C Bowler¹, Dua B. Azhar¹, Cambria M Jensen¹, Hyun-Woo Lee¹, James G Heys¹; ¹University of Utah

Presenter: John C Bowler

We learn by adapting and reusing elements of prior experience, suggesting a key role of curriculum in shaping skills for novel problems. However, it remains unclear how the knowledge for complex tasks is built into neural networks for later reuse. This challenge is compounded by the fact that complex tasks unfold over specific timescales, requiring actions to be suppressed or initiated depending on context and sensory input. Using Recurrent Neural Network (RNN) models trained on a context-dependent timing task, we identify distinct experience-dependent strategies, and find that curriculum-trained RNNs outperform those trained directly on the full task, especially under increasing noise. We then identify signatures of prior experience in both network connectivity and activity. Finally, using Neuropixels recordings from the Medial Entorhinal Cortex (MEC) of mice, we show that these dynamics align with those observed in animals trained on the same curriculum.

Topic Area: Memory, Learning & Knowledge Structures