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

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Interpreting Neural Encoding Models: Voxel-Level Feature Discovery with Causal Validation

Idan Daniel Grosbard¹, Mor Geva¹, Galit Yovel¹; ¹Tel Aviv University

Presenter: Idan Daniel Grosbard

Deep neural network-based neural encoders have advanced our understanding of human visual system. However, most works have relied on correlational methods to produce hypotheses about functional selectivity, with limited causal validation. In this work, we trained an interpretable neural predictor and localized the representations used to predict each voxel's activation for individual images. We then used Large Language Models (LLMs) to decode these representations and perform counterfactual analyses to empirically examine the faithfulness of these descriptions. Overall, this work enables discovery of image-specific critical features for millimeter-resolution brain regions, allowing for a finer-grained understanding of neural response patterns.

Topic Area: Methods, Tools, Theory & Neural Coding