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

Poster D in Poster Session D: Wednesday, August 5, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

Pooling Across Brains: Embedding‑SRM Enables Encoding Models in Ultra‑Low‑Data ECoG

ARNAB BHATTACHARJEE¹, Bobbi Aubrey¹, Werner K Doyle², Patricia Dugan³, Daniel Friedman⁴, Orrin Devinsky⁴, Adeen Flinker², Peter Ramadge¹, Uri Hasson⁵, Samuel Nastase⁶; ¹Princeton University, ²New York University, ³NYU Grossman School of Medicine, ⁴NYU Langone, ⁵Weizmann Institute of Science, ⁶University of Southern California

Presenter: ARNAB BHATTACHARJEE

Large language models (LLMs) have emerged as models of neural activity during naturalistic language processing, using subject‑specific linear encoding models to map contextual embeddings onto individual brain responses. However, these approaches require large amounts of per‑subject data, which is often infeasible in ECoG. Recent work shows that functional alignment via a shared response model (SRM) improves encoding and supports cross‑subject generalization, but existing SRM methods require participants to experience the same stimulus. We introduce Embedding‑SRM (eSRM), which applies SRM to subjects’ encoding weight matrices rather than neural responses, eliminating the need for shared input and enabling alignment across disjoint recordings. We developed a framework using eSRM that allows us to pool neural data from other subjects to effectively enlarge the training set for fitting encoding models in a low-data target participant. This yields substantial gains in encoding performance, particularly when data are extremely limited.

Topic Area: Methods, Tools, Theory & Neural Coding