Beyond the Average Viewer
Predicting individual brain-response differences with TRIBE v2
What this asks
People watching the same movie share most of their brain response, and each person also deviates a little. This project isolates that deviation, learns it from three movie-fMRI runs, and predicts it on an untouched fourth run holding different segments.
The question is about representation, not about a new model. TRIBE v2 is a frozen encoder trained to predict brain responses to a movie, and its 2,048 features are computed from the 11,008 video, audio and text features it is given. We drop its own readout and fit our own decoder from each representation on equal terms, so what changes between them is how easily one person's deviation can be read out.
Result
On a run held out from fitting, your own decoder beats a stranger's, and TRIBE v2 strengthens the effect.
- TRIBE v2
- 0.0424
- its input features
- 0.0201
- difference, 95% CI 0.0183 to 0.0265
- +0.0223
Participant specificity: accuracy with your own decoder minus accuracy with a stranger's, as family-equal means over 174 Human Connectome Project participants in 89 family units. The difference is positive in all four held-out runs and in 80 of those 89 family units. Every participant still needs their own movie fMRI to calibrate.
Poster
Beyond the Average Viewer (PDF, 0.4 MB)
Stanford Bio-X symposium poster. The frozen checkpoint is from d'Ascoli et al., 2026 (arXiv:2605.04326).
Status
The manuscript is in preparation. The analysis code and the participant-free aggregate results will be released with it, under an archival DOI so the cited version stays fixed. Nothing here redistributes Human Connectome Project participant data, family identifiers, movie stimuli or model weights.
Questions, or want to know when the paper lands? rhsr@stanford.edu