Researchers study how complex models predict brain responses to movies
New findings show how pretrained visual features and brain history affect the accuracy of predicting blood-oxygen-level-dependent responses.
Complex neural models predict blood-oxygen-level-dependent (BOLD) responses to naturalistic stimuli. The predictive value of adding complexity remains unclear.
Researchers compared linear diagnostic references and nonlinear predictors. They used HCP movie fMRI with 154 subjects, BOLD Moments with 10 subjects, and Algonauts 2025 with 4 subjects.
Pretrained visual features generally improve prediction over raw pixels. Measured BOLD history provides substantial gains, but these gains generally diminish as rollouts lengthen.
A BOLD response is a blood-oxygen-level-dependent response. A rollout is a recursive prediction without additional measured BOLD. A linear autoregression is a history-conditioned baseline.
Regional response consistency across subjects is associated with held-out prediction accuracy. Less consistent regions tend to gain more from measured history.
Slower hemodynamic responses generally sustain high temporal prediction correlations over longer rollouts. These findings support evaluating stimulus representations and downstream complexity separately.
Key facts
- Pretrained visual features improve prediction over raw pixels.
- Measured BOLD history provides substantial gains that diminish as rollouts lengthen.
- Regional response consistency across subjects is associated with held-out prediction accuracy.