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

Interpretable Language-Model Encoding of Voxel-Level Tuning During Language Comprehension

Michael A. Lepori¹, Kendrick Kay², Greta Tuckute³; ¹Brown University, ²University of Minnesota, ³Harvard University

Presenter: Michael A. Lepori

What are the fine-grained neural population response profiles that support language comprehension? Using a high-field 7T fMRI dataset from eight participants listening to 200 diverse sentences, we leverage recent advances in mechanistic interpretability to develop encoding models that explain a voxel's response by linear combinations of interpretable, small feature sets. Our contributions are three-fold: (1) We recover previous interpretations of voxel populations tuned to either predictability or content, (2) we identify and interpret reliable, participant-specific voxel populations, and (3) show that the human language network shares a common set of features across its constituent regions, with finer-grained variation across individuals. Together, these results provide a framework for studying and interpreting how linguistic information is organized within and across individual brains.

Topic Area: Auditory, Speech & Language Processing