Source-linked AI summary
From Activation to Specificity: Automating Counterfactual Testing of Visual Representations in the Human Brain
Yuval Golbari, Navve Wasserman, Matias Cosarinsky, Roman Beliy, Aude Oliva, Antonio Torralba, Michal Irani, Tamar Rott Shaham
TL;DR
BrainTRACE addresses the difficulty of distinguishing genuine visual concept representations from neural responses to correlated cues. It generates controlled positive, counterfactual, and semantic-negative stimuli, predicts fMRI responses, and tests candidate regions for specificity. Across hundreds of concepts, it recovers known localizations while showing that activation-only methods produce many false positives.
Problem
Existing localization methods do not establish whether strong responses reflect the target concept or correlated visual and semantic cues.
Method
BrainTRACE combines generative models, language-model counter-hypotheses, image-to-fMRI encoding, and counterfactual-specificity scores to identify selective candidate representations.
Results
Over 70% of activation-based localizations failed counterfactual-specificity evaluation, while BrainTRACE recovered known regions across 260 visual concepts and identified additional candidates.
Takeaways & Limitations
Counterfactual-specificity testing makes visual concept mappings more reliable and interpretable while supporting automated follow-up experiment design.
Takeaways & Limitations
BrainTRACE depends on current language and vision models, whose failures include generating semantically similar alternatives that do not reliably exclude the target concept.
Abstract
from arXiv · showhide
Identifying which brain regions represent a visual concept in the human brain is a central challenge in neuroscience. Existing approaches have localized coarse functional regions (e.g., faces, places) through activation maximization, identifying regions that activate strongly for a target concept relative to other concepts. Yet strong activation alone does not establish that a region represents the concept itself, as responses may instead be driven by correlated visual or semantic cues. We introduce BrainTRACE (Testing Representations through Counterfactual Evidence), an automated framework that combines generative and brain models to synthesize controlled stimuli and validate neural representations through targeted counterfactual-specificity testing. Given a query specifying a concept of interest, our framework constructs targeted stimulus sets comprising concept images, counterfactual edits that remove the target concept while preserving other image content, and images with candidate correlated distractors. It then uses an image-to-fMRI encoding model to predict brain responses and searches for representations that respond specifically to the target concept over correlated alternatives. BrainTRACE returns validated candidate representations and proposes follow-up fMRI experiments to further test or extend its discoveries. Our approach successfully recovers known functional localizations and identifies new candidate representations across dozens of concepts, validated on both predicted and measured fMRI data. Critically, we show that without counterfactual evaluation, a large fraction of localizations would be false positives, confirming that activation alone is insufficient evidence of representation.
1 Introduction
BrainTRACE addresses the problem that strong neural activation can reflect correlated visual or semantic cues rather than genuine concept representations. It uses counterfactual-specificity testing to discover, validate, and experimentally extend concept-specific brain representations.
- Existing methods may mistake responses driven by color, background, pose, or other co-occurring cues for representations of the target concept.
- BrainTRACE constructs targeted stimuli, predicts brain responses with an image-to-fMRI model, and identifies voxels or regions selective for the target concept over correlated alternatives.
- The framework characterizes candidate representations by their specificity, anatomical location, confidence, and support from measured data, while proposing follow-up stimuli when evidence is insufficient.
- Over 70% of representations identified by activation-based methods failed counterfactual-specificity evaluation despite high activation scores.
- Across 260 visual concepts, BrainTRACE recovered known face, body, place, and word-selective regions and identified broader fine-grained candidate representations.
- Counterfactual-specificity testing provides a unified framework for discovering, validating, and refining visual representations while identifying gaps and informative follow-up stimuli.
2 Related Works
Visual neuroscience has progressed from low-level maps and coarse category regions toward large-scale concept analysis. BrainTRACE combines generative image control, language-model counter-hypotheses, and image-to-fMRI encoding to test concept specificity and guide experiments.
- Early fMRI research established retinotopic organization and representations of basic properties such as orientation, spatial frequency, color, and motion.
- Encoding models enable voxel-wise fMRI response predictions for images not measured in the scanner, supporting larger-scale concept analysis.
- Modern generative models synthesize and edit images with fine-grained semantic control, while language models propose correlated counter-hypotheses and informative interventions.
- BrainTRACE combines these capabilities with image-to-fMRI encoding to construct controlled stimuli that disentangle concept-driven activation from correlated factors.
3 The BrainTRACE Framework
BrainTRACE builds concept-targeted stimulus sets, scores voxel specificity against semantic and counterfactual alternatives, and combines predicted and measured-data evidence to select candidate representations. It also accounts for measured-data coverage when assigning confidence and designing follow-up experiments.
- BrainTRACE takes a target concept and subject image–fMRI dataset as input, producing a candidate voxel set and confidence estimate.
- The stimulus set includes positive images, semantic negatives representing correlated alternatives, and counterfactual edits that remove the target concept while preserving other content.
- Retrieved positive and semantic-negative images from measured datasets enable evaluation for concepts and alternatives sufficiently represented in the fMRI data.
- Each voxel receives an activation score plus specificity scores comparing positives with hardest semantic negatives and hardest edited counterparts.
- Candidate regions are formed from voxels with positive or highest specificity scores and evaluated on held-out generated data and measured fMRI data.
- Final decisions combine specificity evidence with measured-data concept coverage, returning confidence and informative stimuli for follow-up experiments when coverage is limited.
4 Results
BrainTRACE evaluates visual-concept representations with counterfactual specificity, reducing false discoveries while preserving activation and recovering both known and fine-grained brain organization.
- Counterfactual-specificity testing: Nearly 70% of activation-based discoveries were false positives under counterfactual-specificity evaluation.These regions had high activation but failed to show concept-specific responses, indicating sensitivity to correlated factors.
- Counterfactual-specificity testing: Specificity-based ranking reduced the false positive rate from 73.4% to 23% and increased the true positive rate from 26.6% to 38.7%.Regions with negative train specificity scores were withheld from discovery.
- Quantitative comparison: BrainTRACE maintained comparable activation while improving semantic-negative specificity from -0.44 to 0.62 on generated data and from 0.27 to 0.71 on measured data.The comparison used regions of 100 voxels for the top 50 concepts across four subjects.
- Known-region validation: Specificity-ranked regions aligned with established functional areas for Bodies, Faces, Places, and Words.Alignment was measured as the fraction of top K voxels within corresponding NSD-localized functional regions for voxel sets of 100, 200, and 500.
- Fine-grained localization: Voxel-wise specificity maps localized concepts in high-level visual cortex, including tools near body-part and action-related regions and animal faces in face-selective areas.Specificity scoring produced more spatially localized high-scoring voxels than activation-based methods.
- Cross-subject consistency: Concept-specific representation locations showed clear correspondence across subjects despite variability in functional organization.The authors interpret this correspondence as evidence for robust aspects of visual-cortex organization.
- Fine-grained localization: Related concepts showed distinct voxel patterns across nearby semantic categories, including body-related and text-related concepts.The reported patterns were observed across face-, body-, word-, and object-related visual areas.
- Analysis and limitations: Semantic-negative generation failures reflected limitations of current language and vision models in excluding the target concept while generating similar alternatives.This analysis appears in Appendix A.7.
5 Conclusions
BrainTRACE emphasizes counterfactual-specificity testing for discovering reliable visual concept representations, while acknowledging that current model and stimulus-generation errors can leave correlated factors untested.
- BrainTRACE supports more specific and reliable concept representations, including fine-grained concepts beyond broad category-level regions.
- The framework remains limited by possible errors in image generation, editing, retrieval, verification, and counterfactual scope.
- Some discovered representations may be specific only to the alternatives considered by the current pipeline.
A Ablation & Analysis
Specificity maps show correspondence across subjects for several concepts, while comparison with activation-based localization indicates that specificity scoring produces more selective concept maps.
- Cross-subject consistency: Specificity maps for Animal, Human Interaction, and Hands in Action show clear correspondence across subjects despite variability in individual responses.
- Specificity versus activation: Activation-based localization produces broad high-response patterns, whereas BrainTRACE specificity scoring yields more selective maps by suppressing correlated visual and semantic cues.
A.3 Analysis of Ranking Scores
Ranking analyses show that activation alone favors strong responses but weaker specificity, whereas complementary semantic, measured-data, and counterfactual signals produce better region quality.
- Max-activation ranking achieves the highest activation scores but performs substantially worse on specificity criteria, especially for generated semantic negatives.
- Retrieval from measured data improves measured activation and specificity relative to retrieval from a large image pool, while filtering further improves specificity.
- SSG achieves the strongest generated specificity, SSM the strongest measured specificity, and CFG by far the best edit score.
- CFG+SSG+SSL+MALF+SSM achieves the highest average score across all criteria, supporting integration of complementary ranking signals.
A.4 Region Size Analysis
Across region sizes from 50 to 1000 voxels, BrainTRACE maintains stronger specificity than activation-based methods, although selectivity generally decreases as regions grow.
- BrainTRACE consistently obtains the strongest specificity scores across region sizes and generated-negative, measured-negative, and counterfactual-edit evaluations.
- Activation-based methods achieve stronger activation on generated positives, while MindSimulator+ often achieves the highest measured activation scores.
- Scores gradually decrease for all methods as region size increases, indicating that larger regions become less selective.
- BrainTRACE's relative specificity advantage remains stable across the tested region sizes.
A.5 Quantitative Results Across Subjects
Across four NSD subjects, BrainTRACE consistently ranked regions with the strongest specificity, while activation-based methods often achieved stronger activation scores. Measured-data validation depends strongly on concept coverage, and semantic-negative generation can still produce false positives.
- Quantitative comparison: BrainTRACE consistently achieved the strongest specificity scores on generated negatives, measured negatives, and counterfactual edits across all four subjects.The comparison used subjects 1, 2, 5, and 7, regions of size 100, and the top 50 concepts selected by each method.
- Quantitative comparison: Activation-based methods, especially Max Activation, produced the strongest generated activation scores, while MindSimulator+ often led measured activation scores.
- Measured-data coverage: Measured-data validation is less informative when a concept has few verified positive examples because sparse coverage provides limited evidence about its representation.Positive-example coverage ranges from very few examples to nearly 200 across concepts.
- Measured-data coverage: BrainTRACE evaluates semantic-negative coverage by retrieving verified examples for target–negative pairs, supporting specificity tests on measured data.Retrieval requires the target concept to be absent and the intended negative concept to be present.
- Failure cases: Some false positives remain, particularly for broad properties such as lighting contrast, sky, and reflection.Generated semantic negatives can visibly retain the target concept, including sky regions or reflective surfaces.
A.8 Statistical Testing of BrainTRACE Discovered Regions
BrainTRACE tests whether discovered regions score higher for their target concepts than for concept-specific baselines using one-sided empirical p-values on five validation criteria. The resulting pass counts differ substantially between generated and measured evaluations.
- Test design: The baseline test fixes each discovered 100-voxel region and compares its target-concept score with scores for distinct concepts selected by an LLM.
- Statistical testing: 160 of 260 concept–subject pairs passed Activation–Generated at p ≤ 0.05, compared with 97 passing Activation–Measured.
- Statistical testing: 173 of 260 pairs passed Semantic Specificity–Generated, while 47 passed Semantic Specificity–Measured at p ≤ 0.05.
- Statistical testing: 101 of 260 concept–subject pairs passed Counterfactual Specificity at p ≤ 0.05.
- Test design: The one-sided empirical p-value tests whether the target concept score exceeds the baseline distribution using held-out validation data after region selection.
B Additional Visual Results
BrainTRACE identifies partially distinct representations for related body-action concepts and overlapping but distinct representations for human and animal concepts across high-level visual cortex.
- Action and body representations: Person Running, Person Jumping, and Hands in Action show partially distinct spatial patterns across high-level visual cortex.Binary maps display voxels with positive specificity evidence alongside matching example images.
- Human and animal representations: Human and Animal representations show overlapping but distinct spatial patterns across high-level visual cortex.Binary maps display voxels with positive specificity evidence and corresponding example images.
C.1 Brain Data
BrainTRACE builds predicted and measured image–fMRI analyses from controlled positive, semantic-negative, and counterfactual stimuli. It scores voxel and region responses for activation and specificity, selects positive-specificity candidate regions, and evaluates them on held-out and measured data.
- Brain and image data: BrainTRACE combines measured image–fMRI data with 120K COCO images whose subject-specific fMRI responses are predicted for large-scale retrieval.
- Brain and image data: The image-to-fMRI encoder predicts subject-specific voxel responses for arbitrary images, while voxels below Pearson correlation 0.2 on held-out data are filtered out.
- Stimulus construction: The stimulus set includes positive images, semantically negative images depicting correlated alternatives, and counterfactual negatives edited to remove the target concept.
- Voxel scoring: The Positive Score averages voxel activation over images depicting the target concept.
- Voxel scoring: Semantic-Negative Score compares positive activation with the top 10 semantic negatives having the highest activation for that voxel.
- Voxel scoring: Counterfactual Score compares each positive image with its hardest edited version to measure specificity under controlled concept removal.
- Voxel scoring: A high counterfactual specificity value indicates that voxel activation decreases when the target concept is removed or replaced while other image content remains similar.
- Region discovery: BrainTRACE selects candidate regions as voxels with positive average semantic-negative and counterfactual specificity, then evaluates region-level scores on separate splits and measured fMRI data.