| name | brain-cause-causal-visual-representations |
| description | BrainCause methodology for discovering and causally validating visual representations in the human brain using generative models, counterfactual stimulus synthesis, and fMRI encoding models. Use when: (1) studying causal vs correlational brain representations, (2) designing controlled fMRI experiments with counterfactual stimuli, (3) validating whether brain regions truly represent specific visual concepts beyond activation-based localization, or (4) applying generative AI to neuroscience brain mapping. |
| arxiv_id | 2605.23895 |
| published | 2026-05-22 |
| authors | Yuval Golbari, Navve Wasserman, Matias Cosarinsky, Roman Beliy, Aude Oliva, Antonio Torralba, Michal Irani, Tamar Rott Shaham |
| tags | ["causal-representation","fmri-encoding","generative-models","counterfactual-stimuli","brain-mapping","visual-neuroscience","functional-localization"] |
BrainCause: Causal Visual Representation Discovery in the Human Brain
Core Concept
BrainCause is an automated framework that combines generative image models with image-to-fMRI encoding models to synthesize controlled stimuli and validate neural representations through targeted causal testing. It moves beyond traditional activation-based functional localization to establish whether brain regions genuinely represent visual concepts rather than merely responding to correlated cues.
Key Insights
-
Activation ≠ Representation: Strong activation alone does not establish that a brain region represents a concept — responses may be driven by correlated visual or semantic cues. Without causal validation, a large fraction of functional localizations would be false positives.
-
Counterfactual Causal Testing: BrainCause constructs three types of controlled stimuli:
- Concept images: Strong exemplars of the target concept
- Counterfactual edits: Images where the target concept is removed while preserving other content
- Correlated distractor images: Images with candidate confounds that correlate with the target concept
-
Image-to-fMRI Encoding: Uses a predictive encoding model to estimate brain responses to synthetic stimuli, enabling large-scale causal testing without requiring new fMRI scans for every hypothesis.
-
Automated Experiment Proposal: Returns validated candidate representations and proposes follow-up fMRI experiments for further testing.
Methodology
Stimulus Construction Pipeline
- Define target visual concept (e.g., "faces", "places", "body parts")
- Generate concept exemplars using generative models
- Create counterfactual versions that remove the target concept
- Generate correlated distractor stimuli
Validation Procedure
- Feed synthetic stimuli through image-to-fMRI encoding model
- Search for voxels/regions responding specifically to target concept
- Control for correlated alternatives via comparison conditions
- Return validated candidate representations
Recovery Validation
- Successfully recovers known functional localizations (e.g., FFA for faces, PPA for places)
- Identifies new candidate representations across dozens of concepts
- Validated on both predicted fMRI (via encoding model) and measured fMRI data
Applications
- Neuroscience: Mapping the causal basis of visual concept representations in human cortex
- fMRI experiment design: Generating hypothesis-driven stimulus sets for targeted validation
- Brain-AI alignment: Understanding which features drive neural responses vs. truly encoded representations
- Cognitive neuroscience: Disentangling correlated visual/semantic confounds from genuine concept selectivity
Activation Keywords
- brain-cause, causal-representation, counterfactual-fmri, functional-localization, activation-vs-representation, generative-fmri, brain-encoding, visual-concept-validation, stimulus-synthesis, fmri-experiment-design