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

Geometric Signatures of Event Boundaries in Brain Foundation Models

Kyungjin Oh¹, Jiook Cha¹; ¹Seoul National University

Presenter: Kyungjin Oh

Brains segment continuous experience into discrete events at moments of prediction error. These boundaries shape what is remembered and how experiences are interpreted. We ask whether a resting-state fMRI foundation model can capture neural event structure during movie watching. Using SwiFT_v2, a pretrained resting-state fMRI foundation model, applied to StudyForrest (N=15, Forrest Gump, ~162 behavioral event boundaries), we find that event boundary information is not decodable per-timepoint but emerges when averaging across many boundaries as a significant increase in embedding distance (p=.0003). These transitions involve representational displacement (cosine, Procrustes distance significant) without restructuring (local dimensionality, effective rank null). Temporal autocorrelation drops at boundaries (p=.028), providing a direct prediction failure signature. Signals appear only in later transformer layers (p<.02), and cross-subject alignment decreases at boundaries (p<.001). Foundation model embeddings thus capture event structure as continuous geometric transitions, complementing HMM's discrete detection.

Topic Area: Decision-Making, Cognitive Control & Event Cognition