BrainExplore: Large-Scale Discovery of Interpretable Visual Representations in the Human Brain

1 Introduction

• High-dimensional fMRI, relatively small image-to-fMRI datasets, and manual analysis have left studies focused on selected concepts or regions, motivating scalable discovery across the cortex.Source: §1 · p. 1
• BrainExplore uses unsupervised, data-driven methods to discover brain-activity patterns and explain them in natural language, drawing inspiration from automated interpretability in vision and language models.Source: §1 · pp. 1–2
• The pipeline decomposes fMRI responses into component patterns, retrieves each pattern’s top-activating responses and images, and selects the best-aligned concept with an alignment score.Source: §1 · p. 2, Figure 1 · p. 2
• Adding predicted fMRI expands the image pool from approximately 10k measured images to over 120k measured or predicted responses; SAE yields many interpretable patterns, including representations complementary to other methods.The authors report that augmentation improves decomposition quality and increases the diversity of discovered representations.Source: §1 · p. 2
• BrainExplore discovers thousands of interpretable patterns across the brain, including fine-grained visual concepts such as object identity, body parts and poses, and indoor and outdoor scenes.The findings are quantitatively evaluated on measured fMRI not used during interpretation.Source: §1 · pp. 2–3
• The authors propose releasing a brain-inspired concept dictionary, a large-scale image–fMRI–explanation ranking dataset, and code as a benchmark for future work.Source: §1 · p. 3

2 Related Works

• Category-selective studies have mapped regions and concepts, but rely on predefined categories, relatively clean images, and separate hypothesis tests, limiting discovery of nuanced or overlapping representations.Source: §2 · p. 3
• Earlier fMRI decomposition studies typically examined one method and a few components through manual inspection; BrainExplore compares methods at scale, expands the image pool with image-to-fMRI models, and introduces SAE decomposition.Source: §2 · p. 3
• Image–fMRI approaches often analyze individual voxels or whole regions, whereas BrainExplore uses predicted responses to augment decomposition and retrieve activating examples at the pattern level.Source: §2 · p. 3
• BrainExplore bridges automated interpretability and fMRI decomposition in a large-scale pipeline that discovers brain patterns and benchmarks decomposition methods.Source: §2 · pp. 3–4

3.1 Experimental Setting

• The evaluation protocol uses NSD measured and predicted fMRI, with decomposition performed separately by ROI and learned from brain responses without image features or labels.The dataset includes eight subjects and about 73,000 measured image–fMRI pairs; about 120,000 predicted responses augment training and image retrieval, while interpretations are verified on measured fMRI.Source: §3.1 · p. 4, §3.2 · p. 5

3.3 From fMRI Patterns to Visual Concepts

• For each pattern, the framework retrieves top-activating images and presents selected pattern explanations, including examples from PPA.Figure 4 shows Subject 1 PPA patterns with their top activating images and selected explanations; PPA is associated with scenes and places.Source: Figure 4 · p. 6

3.4 Scaling to an Unlimited Number of Patterns

• To scale beyond explaining each pattern individually, BrainExplore builds a brain-inspired concept dictionary and labels images against its candidate concepts.It selects coherent patterns to generate explanations, consolidates these into about 1,300 concepts, then uses CLIP shortlisting and two-pass VLM verification to produce image–hypothesis labels.Source: §3.4 · p. 6

3.5 Discover Interpretable Brain Patterns

• The Discover step scores each pattern against candidate concepts using the consistency of concept labels among top-activating images, with a rarity adjustment and an average across measured and predicted pools.It supports searches for the most interpretable patterns or the best-explained pattern for a chosen hypothesis, within an ROI or across ROIs.Source: §3.5 · pp. 6–7

4.2 Quantitative Evaluation

• At the 0.5 threshold, SAE yields approximately 9k interpretable patterns across the brain, compared with ∼6k for Voxels and 226 for ICA.Near-duplicate patterns are excluded when their voxel-wise correlation with a higher-scoring pattern exceeds 0.5.Source: §4.2 · p. 7
• With predicted fMRI, ICA rises from around 12 to 207 interpretable hypotheses at the 0.5 threshold, while SAE rises from around 105 to 218.Predicted responses are used both to learn decompositions and to retrieve top-activating images; SAE yields the most interpretable hypotheses with or without predicted fMRI.Source: §4.2 · p. 7
• Combining ICA and SAE yields the strongest performance, enabled by automatic scoring that supports aggregation across decompositions.Evaluation assigns best pattern–hypothesis pairs on ranking halves and reports metrics on disjoint evaluation halves; the main paper reports Subject 1 results and appendices cover other subjects.Source: §4.1 · p. 7, §4.2 · p. 8

4.3 Alignment with Established Neuroscientific Knowledge

• Discovered concepts align with established regional selectivity, and interpretable concepts show high but imperfect consistency across subjects.Examples include EBA with actions and sports, FFA with people and faces, place regions with locations, and text-related concepts with OWFA and VWFA.Source: §4.3 · p. 8

4.4 Visual Representations in the Human Brain

• BrainExplore finds fine-grained representations, including sports and actions in EBA and distinct landscape, commercial-building, and stone-architecture concepts in PPA.Different ROIs best explain different concepts; SAE patterns are more compact and clustered than ICA patterns despite receiving no spatial information or constraint.Source: §4.4 · pp. 8–9

4.5 Ablations & Analysis

• The appendices analyze SAE sparsity and dimensional expansion, predicted-fMRI pool size, per-ROI metrics, and complementarity across decompositions.Source: §4.5 · p. 9

5 Discussion & Limitations

• VLM-based labeling may be noisy, and definitive validation of newly identified representations requires future targeted neuroscience experiments.The authors use one-concept-at-a-time labeling followed by a second VLM verification step to mitigate labeling noise; they expect the framework to help guide experiment design.Source: §5 · p. 9

SAE ablations

• Higher sparsity generally yields more interpretable SAE patterns, and larger expansion factors tend to perform better, with 0.5 performing noticeably worse in most configurations.Expansion factors above 4 were computationally expensive without consistent additional gains; sparsity coefficients above 4 showed no clear improvement.Source: A Ablations and Additional Results · p. 14, Table T1 · p. 14

Predicted fMRI pool size

• Enriching the retrieval pool with predicted fMRI improves interpretability over measured-fMRI-only pools, and the largest pool gives the best-performing model its highest, most stable scores.Training data are fixed across these comparisons; larger pools also increase the number of images used to compute hypothesis scores, making scores more robust.Source: A Ablations and Additional Results · p. 14, Table T2 · p. 15

Per-ROI quantitative results

• Higher-level visual regions show greater interpretability than early visual areas in the per-ROI results.The results report the percentage of explained hypotheses above 0.5 on a pool enriched with all predicted fMRI; V1–V3 and FBA, FFA, and VWFA are averaged across their specified partitions or subareas.Source: A Ablations and Additional Results · p. 14, Table T3 · p. 15

Complementarity across decompositions

• SAE and ICA are the most complementary decomposition pair, each substantially improving the other when their patterns are combined.The reported gain is the increase in explained hypotheses over the better of the two methods used separately.Source: A Ablations and Additional Results · p. 15, Table T4 · p. 15

B.1 Alignment between ROI-discovered representations and known functionality

• Across subjects and ROIs, functionally aligned concepts often receive pipeline scores above 0.5; ROIs with fewer than five relevant concepts are excluded.Concepts are restricted to those interpretable in at least three subjects, and the reported ROI score is the fraction of relevant concepts also interpretable in that subject and ROI.Source: §B.1 · p. 16, Figure S1 · p. 16

B.2 Cross-subject consistency of interpretable concepts

• Cross-subject concept consistency is measured by IoU across subject subsets of size 2, 4, and 8, with ROI-randomization permutations used to test whether overlap exceeds chance.IoU is the intersection size divided by the union size of the corresponding concept sets.Source: §B.2 · p. 17

B.3 Cross-subject consistency of interpretable patterns

• For each Subject 1 SAE pattern, the analysis finds its highest-IoU counterpart in each target ROI and averages best-match scores by source ROI.Similarity is the IoU between sets of top-activating images; the Subject-1-to-Subject-1 diagonal is omitted from the heatmaps because its within-ROI values are especially high.Source: §B.3 · p. 17
• SAE-pattern similarity structure is broadly consistent across subjects, and functionally related ROIs tend to show higher mutual similarity.Early visual ventral regions cluster more strongly with one another than with dorsal regions; absolute IoU values are modest because identical top-activating images are unlikely.Source: §B.3 · p. 17, Figure S2 · p. 18

C.1 SAE Training Details

• The SAE is fully linear, applies ReLU after the encoder, and trains with reconstruction MSE plus an ℓ1 penalty on latent activations.The expansion factor sets latent dimensionality relative to input size, while the sparsity coefficient weights the sparsity loss against reconstruction loss.Source: §C.1 · p. 19
• Measured and predicted fMRI use separate encoder networks but a shared decoder, allowing both data types to map to the same dictionary patterns.Each training batch contains equal numbers of measured and predicted samples to balance gradients.Source: §C.1 · p. 19
• The main hyperparameter search tests sparsity coefficients and expansion factors of 1, 2, and 4; the single-model configuration uses 4 for both.The multiple-model setup uses four configurations selected by rank score on Subject 1, and the same hyperparameters are applied across subjects.Source: §C.1 · p. 19

C.2 Baseline Decomposition Training Details

• Baselines used measured-only fMRI (∼10k samples) or combined measured and predicted fMRI (∼120k samples), with component and sign conventions adapted to each decomposition.Voxels, PCA, ICA, and NMF were included; ICA and NMF used component counts tied to PCA explained-variance configurations.Source: §C.2 · p. 19
• NMF initialization and regularization had minimal impact on final interpretability, consistent with stable convergence under nonnegativity constraints.Source: §C.2 · p. 20

C.3 Visualize & Explain

• Each pattern is visualized with its six most activating measured-response images and ten predicted-response images, then captions are used to generate 3–12 hypotheses shared by at least half the images.Separate VLM captioning and LLM hypothesis generation because current vision–language models often produce incorrect or overly general cross-image hypotheses directly.Source: §C.3 · p. 20

C.4 Scaling to an Unlimited Number of Patterns

• A hypothesis dictionary and image–hypothesis labels let new patterns be evaluated by checking how consistently their top-activating images express dictionary concepts.Source: §C.4 · p. 20

C.5 Region of Interest (ROI)

• The analysis spans 19 predefined visual ROIs, from early areas associated with basic visual attributes to higher-level regions selective for bodies, faces, scenes, places, words, and object forms.Source: §C.5 · p. 20

C.6 Compuational Resourences

• Hypothesis generation and image labeling each required around 100 GPU hours; running the full pipeline across decompositions and concepts took around four hours.Source: §C.6 · p. 20

• Table T5 counts patterns whose best-hypothesis alignment exceeds 0.5 or 0.8, excluding components correlated above 0.5 with a higher-scoring pattern.Source: Table T5 · p. 21
• Tables T6 and T7 report average interpretable-hypothesis counts and percentages across subjects, respectively, with standard errors.Source: Table T6 · p. 21, Table T7 · p. 21
• Table T8 reports interpretable hypotheses across all subjects as counts and percentages.Source: Table T8 · p. 22

• SAE patterns are more compact, clustered, and spatially localized than ICA patterns despite both methods using one-dimensional voxel vectors without spatial information.The maps use identical preprocessing, color scaling, and selection criteria.Source: E Additional Visualizations · p. 23, Figure S3 · p. 23
• Additional Subject 1 patterns are shown for EBA, RSC, OPA, FBA-1, and FBA-2 with top-activating images and selected explanations.Source: E Additional Visualizations · p. 23, Figure S4 · p. 24, Figure S5 · p. 24
• Additional patterns and concepts are visualized for Subjects 2 and 5, with examples spanning OPA, VWFA-1, EBA, FFA-1, FBA, and hV4.Source: E Additional Visualizations · p. 23, Figure S7 · p. 25, Figure S8 · p. 26
• Full grids show the top 16 activating images for EBA and PPA concepts; EBA grids distinguish predicted and measured responses and order images by activation strength.Source: E Additional Visualizations · p. 23, Figure S11 · p. 28, Figure S12 · p. 29
• ROI concept visualizations show both concepts represented across multiple regions and each concept’s highest-alignment ROI, with alignment scores above 0.5 displayed.Source: E Additional Visualizations · p. 23, Figure S15 · p. 32, Figure S16 · p. 33
• Pipeline prompts depict detailed per-image captioning, shared-hypothesis generation, and image–hypothesis labeling as a separate upscaling step.Source: E Additional Visualizations · p. 23, Figure S17 · p. 34, Figure S18 · p. 34
• Example images are paired with the hypotheses labeled positive for each, drawn from the brain-inspired hypothesis dictionary and predicted-fMRI image set.Source: E Additional Visualizations · p. 23, Figure S20 · p. 35