BrainCause Framework Identifies Visual Representations via Causal Testing
- •MIT researchers launched BrainCause to identify visual concept representations in the human brain.
- •The study proves that neural activation alone is insufficient and causes significant false positives.
- •BrainCause uses generative and brain models to validate neural representations through targeted causal testing.
MIT researchers introduced BrainCause, an automated framework designed to identify visual concept representations in the human brain by moving beyond mere activation metrics. While traditional neuroscience identifies functional regions like faces or places via activation maximization (a method finding regions strongly responsive to a concept), this study confirms such activation is insufficient to prove true representation. Responses often stem from correlated visual or semantic cues rather than the concept itself, leading to significant false positives in conventional mapping studies.
The BrainCause framework integrates generative and brain models to conduct targeted causal testing. For a given query, the system generates controlled stimulus sets, including images of the target concept, counterfactual edits that isolate the concept by removing it while maintaining other content, and sets featuring candidate correlated distractors. By employing an image-to-fMRI encoding model, the researchers predict brain responses to isolate specific representations from alternatives. The method has successfully recovered known functional localizations and identified dozens of new candidate representations. Validation on both predicted and measured fMRI data demonstrates that causal testing is essential for filtering out false positives, proving that neural activation alone provides incomplete evidence for concept localization in the human brain.