acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

Understanding how the human brain represents visual concepts, and in which brain regions these representations are encoded, remains a long-standing challenge. Decades of work have advanced our understanding of visual representations, yet brain signals remain large and complex, and the space of possible visual concepts is vast. As a result, most studies remain small-scale, rely on manual inspection, focus on specific regions and concepts, and rarely include systematic validation. We present a large-scale, automated framework for discovering and explaining visual representations across the human cortex. Our method comprises two main stages. First, we *discover* candidate interpretable patterns in fMRI activity through unsupervised, data-driven decomposition methods. Next, we *explain* each pattern by identifying the set of natural images that most strongly elicit it and generating a natural-language description of their shared visual meaning. To scale this process, we introduce an automated pipeline that tests multiple candidate explanations, assigns reliability scores, and selects the best description for each voxel pattern. Our framework reveals thousands of interpretable patterns spanning many distinct visual concepts, including fine-grained representations beyond well-established categories.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Neuroscience, brain representations, unsupervised representation learning, sparse autoencoders, fMRI encoding

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_mJJLOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Authors are introducing BrainExplore, a framework for automated interpretable pattern discovery given fMRI activity during a passive viewing task. They work on the NSD dataset and propose a multi-stage approach, the fMRI activity within each ROI is decomposed via PCA, ICA or SAE into a set on key components, then linearly combined to estimate the real per-image activity. They took the images most active for each pattern and pass them to VLM to formulate pattern hypothesis and LLMs to validate them looking at what is shared across these images.

SAE gives the highest number of intepretable patterns that cross the agreement threshold and they also use universal brain encoder to augment the amount of data available by a large factor (x12).

The main contribution is how they automatize everything for large scale pattern discovery, which could be an useful tool for neuroscientific hypothesis generation

[EDIT] Given the substantial additional experimental evidence, I am inclined to revise my initial rating upward, while retaining some reservations about the interpretation of the results as novel neuroscientific discoveries.

Critical strengths

The study is conducted well, every passage is clear.

The depth of the appendix is remarkable and honestly anticipated many of questions (cross-subject map ROI explanation agreement, cross-dataset validation, ablation studies and impact of dimensionality of each technique used to decompose the signal).

The scale and the systematic comparison of decomposition methods under a common, automatic score are useful.

Discussion is honest about results and limitations.

Critical weaknesses

I think there are some weaknesses in this work, mainly linked to novelty and use of predicted fMRI data.

  1. The first one about novelty is that the idea of exploring/finding brain patterns on this dataset was already explored by a bunch of other works like BrainDIVE, BrianScuba, "Bridging Brains and Concepts: Interpretable Visual Decoding" and "Evidence for compositionality in fMRI visual representations via Brain Algebra" published in 2024-2025, and probably more. Of course, the method is different and defensible, but at least I think it's important to mention these works and explain why the author think this work is adding something. Please don't take it wrong, but my general impression is that this is just another way to do the same thing, but I'm happy to revise my opinion if authors can provide evidence of why their approach is better. In particular, I'd also like to see, when possible a comparison between patterns discovered in these and other works together with the patterns discovered by BrainExplore. This would strenghten the work a lot in my point of view.

  2. What encoders can learn is in principle all but noise (except for a general noise structure). I'm not convinced by the results of fMRI predicted patterns. Most of these models were trained on the same NSD dataset, and it's hard to say that the benefits come from more syntehtic data rather than some form of denoising/data circularity. I trust more the results (even if more modest) on the original data themselves. I would narrow down each claim on the synthetic part.

2a. How much of the brain structure comes from the encoder? Big jump can be seen fro any decomposition technique when you use it, but I think you are mostly capturing the feature geometry of the original image backbone projected in the brain space rather than true brain activity.

  1. The averaged score can be carried by the predicted pool. The final score averages measured and predicted pools, so a concept scoring 0.9 on predicted and 0.1 on measured passes the 0.5 threshold. The claim that “all interpretations are evaluated and verified on measured fMRI” needs direct support. The authors should report, for models trained with augmentation, the measured-pool-only score on the evaluation half, and the fraction of headline concepts passing on measured data alone.
Questions for authors
  1. The Beliy et al. encoder was trained on NSD responses from the same subjects. Were the evaluation-half images excluded from its training, or could the encoder have seen the responses later used to score patterns? More broadly, how do you separate the benefit of “moore data” from the encoder simply denoising responses toward its own learned structure?

  2. Could you run the same pipeline on the encoder’s image-backbone features directly, with no brain readout and matched dimensionality, or on an encoder fit to permuted fMRI targets? If those yield a concept set comparable to the predicted-fMRI condition, how much of the large gain (e.g., ICA from 12 to >200 hypotheses) reflects brain activity rather than the backbone’s feature geometry projected into voxel space?

  3. For decomposition trained on fMRI predicted, could you report score omputed only on the measured evaluation pool, without averaging with the predicted pool?

  4. Without retraining (You have the re-training example on 1 subject of fMRI LAION) how similar are the patterns you find on NSD vs the one you can find on another dataset? fMRI-LAOIN works well here, or you can use BOLD5000 or GOD or Deeprecon or THINGS-fMRI. If you pick a pattern and look at the images that maximize that specific pattern on a different subject/different dataset, will they be similar? This would be the strongest leverage to claim you can identify some new pattern in the brain.

7 comments

We thank the reviewer for their time, effort, and constructive feedback. We conducted additional experiments addressing their suggestions and responded to each point below. Full experimental details and results are provided in the temporary Appendix J for the reviewer's convenience and will be integrated into the relevant sections of the final paper.


W1. The idea of exploring/finding brain patterns on this dataset was already explored by a bunch of other works like BrainDIVE, BrianScuba, "Bridging Brains and Concepts: Interpretable Visual Decoding" and "Evidence for compositionality in fMRI visual representations via Brain Algebra"[...] I think it's important to mention these works and explain why the author think this work is adding something […] I'd also like to see, when possible a comparison between patterns discovered in these and other works together with the patterns discovered by BrainExplore.

Response. We thank the reviewer for highlighting these relevant works. We agree that the distinctions between BrainExplore and previous approaches were not sufficiently clear and will expand the related work and clarify our contributions throughout the paper. BrainExplore makes four main contributions: (i) Distributed representation discovery. We extend previous work on distributed fMRI representations into a large-scale automated discovery and interpretation framework, systematically exploring fine-grained patterns across brain regions and decomposition methods, rather than focusing on individual voxels or broad predefined ROIs. (ii) SAE decomposition. We introduce sparse autoencoders (SAEs) for fMRI decomposition, demonstrating their ability to recover substantially more interpretable patterns, including representations complementary to standard decomposition methods. (iii) Scalable Exploration. Our framework builds a brain-inspired dictionary of ~1,300 concepts and precomputes image–concept labels, enabling efficient evaluation of hundreds of thousands of patterns without costly VLM/LLM explanations for each, as required by previous approaches. (iv) Extensive evaluation and validation. We provide more extensive systematic evaluation, quantitatively comparing discovered representations across decomposition methods, brain regions, and subjects, alongside validation on held-out measured fMRI, alignment with established neuroscience, and generalization across encoders and datasets (e.g. LAION-fMRI).

In comparison, BrainSCUBA and LaVCa focus on individual voxels, with LaVCa scalability limited by per-voxel VLM/LLM interpretation. BrainDiVE explores predefined functional regions through image synthesis limiting the scope and scale of systematic discovery. Bridging Brains and Concepts starts with image-defined concepts and learns to predict them from fMRI through a decoder, rather than discovering representations directly from brain activity. Brain Algebra investigates compositionality using patterns constructed for only 12 predefined concepts. Both primarily focus their evaluation on image reconstruction rather than systematically validating the semantic selectivity of brain representations on held-out measured fMRI. In contrast, BrainExplore combines SAE-based decomposition with scalable interpretation and extensive quantitative validation, enabling the discovery and evaluation of substantially more brain representations.

New quantitative comparison. To further address the reviewer's request, we conducted an additional comparison with two methods for which it was feasible: BrainSCUBA and LaVCa, using NSD Subject 1 across EBA, FFA, and PPA. We mapped their voxel captions to our concept dictionary and evaluated all methods using the same held-out measured fMRI scoring procedure. BrainExplore (SAE+ICA) recovered 239/62 concepts at alignment thresholds of 0.5/0.8, compared with 86/12 for direct voxel search, 51/10 for LaVCa, and 31/1 for BrainSCUBA. While this comparison is approximate due to methodological differences and necessary implementation adaptations, it provides additional evidence of broader concept recovery by BrainExplore.

W2. What encoders can learn is in principle all but noise (except for a general noise structure). I'm not convinced by the results of fMRI predicted patterns. Most of these models were trained on the same NSD dataset, and it's hard to say that the benefits come from more syntehtic data rather than some form of denoising/data circularity. I trust more the results (even if more modest) on the original data themselves. I would narrow down each claim on the synthetic part.

Response. We agree on the importance of validating discovered representations against measured brain activity, and we will clarify this distinction and narrow our claims accordingly. A key motivation for using predicted fMRI is that many fine-grained visual concepts are represented by very few images in the measured dataset, making reliable interpretability scoring difficult. However, a strong multi-subject image-to-fMRI encoder can leverage a much broader training distribution to predict brain responses for additional images containing these concepts. We therefore view representations discovered primarily through predicted fMRI as promising hypotheses for targeted follow-up experiments, rather than already strongly validated neural representations and will better clarify that.

To directly assess the support from measured fMRI, we conducted the additional measured-only evaluation suggested by the reviewer (W3/Q3). This analysis identifies interpretations supported by measured responses alone and distinguishes additional discoveries that lack measured-data support from those that remain inconclusive due to limited evaluation evidence.


W2.a/Q2. How much of the brain structure comes from the encoder? [...] Could you run the same pipeline on the encoder’s image-backbone features directly, with no brain readout and matched dimensionality, or on an encoder fit to permuted fMRI targets?

Response. To directly address the reviewer's concern, we conducted an additional control experiment. We trained two encoders with the same pretrained image backbone and brain-specific architecture, one using the original image–fMRI pairs and the other using randomly permuted correspondences. We then independently repeated SAE discovery using predictions for the same external images across three ROIs of NSD Subject 1. Shuffling reduced the number of discovered concepts from 247 to only 5 at the 0.5 threshold, and from 36 to 0 at the 0.8 threshold. Importantly, when evaluated on held-out measured fMRI alone, the corresponding numbers dropped from 110 to 1 and from 18 to 0.

These findings are further supported by several complementary analyses. The image-to-fMRI encoder used in BrainExplore has demonstrated strong generalization to unseen images (Beliy et al., 2024), supporting its ability to predict meaningful brain responses. Our analysis using alternative image-to-fMRI encoders (Appendix F.2) shows substantial consistency in discovered concepts across different architectures and image backbones. Furthermore, our randomized-activity controls (Appendix B.7) demonstrate that discoveries depend on structured brain activity. Together, these results provide strong evidence that our discoveries rely on learned image–brain relationships rather than simply arising from the pretrained image backbone. Additional experiments evaluate discoveries using measured fMRI alone (W3/Q3), and demonstrate cross-dataset concept consistency (Q4).

W3/Q3. The averaged score can be carried by the predicted pool. The final score averages measured and predicted pools, so a concept scoring 0.9 on predicted and 0.1 on measured passes the 0.5 threshold. The claim that “all interpretations are evaluated and verified on measured fMRI” needs direct support. The authors should report, for models trained with augmentation, the measured-pool-only score on the evaluation half, and the fraction of headline concepts passing on measured data alone. For decomposition trained on fMRI predicted, could you report score computed only on the measured evaluation pool, without averaging with the predicted pool?

Response. We thank the reviewer for raising this concern. To assess the measured-data support for interpretations discovered through augmentation, we repeated the original BrainExplore scoring procedure using only the measured evaluation pool, across all subjects and decomposition methods. We report the number of hypotheses passing the 0.5 and 0.8 thresholds for measured-only data in the Supported columns of the table below. We further examined hypotheses that passed the original combined measured and predicted score but did not pass the corresponding measured-only threshold. Importantly, failure to pass the measured-only threshold does not necessarily mean that a hypothesis is unsupported: the measured evaluation pool may contain too few examples of a particular concept to reliably assess its alignment.

To account for this, we introduced a minimum coverage requirement of 100 concept-positive measured images. Hypotheses below this requirement were classified as Insufficient coverage. For the remaining hypotheses, we assessed uncertainty using 5,000 paired image-level bootstrap resamples with multiple-comparison correction. We classified hypotheses as Not supported when the adjusted 95% upper confidence bound of the measured-only score fell below the corresponding threshold, or as Statistical uncertainty when the evidence was insufficient to establish this. The results are summarized in the table below, with full experimental details provided in Appendix J.8. A substantial amount of hypotheses pass measured-only evaluation, while relatively few additional discoveries are classified as not supported. Most remain inconclusive due to limited coverage or statistical uncertainty.

MethodSupported (>0.5)Not supported (>0.5)Insufficient coverage (>0.5)Statistical uncertainty (>0.5)Supported (>0.8)Not supported (>0.8)Insufficient coverage (>0.8)Statistical uncertainty (>0.8)
Voxels5523328009
PCA5442259004
NMF4420207004
ICA100171479161019
SAE13671567330118
SAE+ICA146112178352122
Reviewer_EF9tOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces BrainExplore, a scalable framework for automatically discovering and explaining interpretable visual representations in human fMRI. The authors combine multiple decomposition methods, including PCA, NMF, ICA, and sparse autoencoders (SAEs), with automated natural-language hypothesis generation and evaluation. Their key contribution is the use of encoder-predicted fMRI to expand the limited measured brain responses, both for learning the decomposition components and retrieving the top-activating images. The framework also provides a common (not novel in these terms) scoring mechanism to evaluate and compare component–concept associations across different decomposition methods.

The authors demonstrate the discovery of 1000s of interpretable patterns (~1300 visual concepts) across visual cortical regions, including fine-grained concepts beyond established visual categories. The results are supported by extensive experiments, including held-out measured-fMRI evaluation, cross-subject consistency, encoder robustness, and independent-dataset generalization. Overall, the main contribution lies in integrating and scaling existing decomposition and automated interpretability techniques for brain representations, rather than introducing automated brain interpretability itself.

Critical strengths
  1. BrainExplore provides a unified framework for automatically generating, evaluating, and retrieving natural-language explanations of distributed fMRI components across multiple decomposition methods. Its reusable hypothesis dictionary and component–concept scoring enable such an analysis at a scale that would be difficult to achieve through manual inspection.

  2. The framework also expands limited measured fMRI data using an image-to-fMRI encoder, integrating the predicted responses into both decomposition learning and retrieval of preferred stimuli (activating images).

  3. Empirical evidence is substantial. This study evaluates discovered patterns on held-out measured fMRI, included semantic-selectivity AUC, matched-component comparisons, cross-subject consistency as well as the human assessment of semantic labels in the Appendix. These additional experiments substantially establish the credibility of the results and inferences.

Critical weaknesses

Major:

  1. The claim that no large-scale automated interpretability pipeline exists for the brain is too broad. Prior works such as BrainSCUBA (2024) and LaVCa (2025) have explored automated natural-language explanations of visual fMRI selectivity, while Brain Dissection (2023) applied automated interpretability techniques to fMRI-trained neural networks. While BrainExplore makes meaningful contributions through pattern-level interpretation, multiple decomposition methods, and extensive validation (as acknowledged in the Strengths section), it does not sufficiently clarify how these contributions differ from closely related prior work. Discussing these approaches, particularly the currently missing Brain Dissection reference, would help establish the paper's novelty more clearly.

  2. Limited validation of concept specificity. Appendix B.6 demonstrated that the explanations generalize to unseen images, but the one-vs-rest ROC-AUC does not establish whether a pattern responds to the assigned concept or correlated visual features. For example, a "surfing" pattern may actually respond to water or body posture. Prior interpretability works, such as FALCON (2023), highlights the importance of contrastive examples for filtering spurious explanations.

Minor:

  1. The spatial localization of SAE patterns without any explicit spatial constraints seems interesting. However, this can arise from correlations between neighboring voxels due to fMRI measurement or preprocessing, rather than distinct neural representations. Could the authos quantify the localization and compare it against spatial-correlation-preserving controls while accounting for sparsity differences across decomposition methods?

References

Brain Dissection: fMRI-trained Networks Reveal Spatial Selectivity in the Processing of Natural Images. NeuRIPS 2023.

FALCON: Identifying Interpretable Subspaces in Image Representations. ICML 2023.

BrainSCUBA: Fine-Grained Natural Language Captions of Visual Cortex Selectivity. ICLR 2024.

LaVCa: LLM-assisted Visual Cortex Captioning. ICLR 2026. ArXiv Feb 2025.

Questions for authors
  1. Could the authors evaluate explanations using matched positive and negative images to distinguish the target concept from correlated visual features? This would provide stronger evidence of fine-grained neural selectivity.

  2. Can you clarify how BrainExplore differs from these existing frameworks and discuss whether direct comparisons are feasible, especially in terms of explanation quality, scalability, and interpretability on held-out fMRI?

  3. (Optional) Can you addresses or clarify the minor weaknesses ?

7 comments

We thank the reviewer for their time, effort, and constructive feedback. We conducted additional experiments addressing their suggestions and responded to each point below. Full experimental details and results are provided in the temporary Appendix J for the reviewer's convenience and will be integrated into the relevant sections of the final paper.


W1/Q2. The claim that no large-scale automated interpretability pipeline exists for the brain is too broad. [...] While BrainExplore makes meaningful contributions through pattern-level interpretation, multiple decomposition methods, and extensive validation (as acknowledged in the Strengths section), it does not sufficiently clarify how these contributions differ from closely related prior work. [...] Can you clarify how BrainExplore differs from these existing frameworks and discuss whether direct comparisons are feasible, especially in terms of explanation quality, scalability, and interpretability on held-out fMRI?

Response. We thank the reviewer for highlighting these relevant works. We agree that automated approaches to brain interpretability have been explored previously, and we will revise the related work and contribution statements accordingly. BrainSCUBA, LaVCa, and Brain Dissection primarily investigate individual voxels, whereas BrainExplore discovers and interprets distributed brain patterns through unsupervised decomposition. Furthermore, unlike Brain Dissection's predefined visual properties, BrainExplore automatically discovers its reusable concept dictionary from brain-response patterns.

We elaborate on our contributions along the three dimensions raised by the reviewer: (i) Explanation quality. Our contribution is not primarily to improve the linguistic quality of captions, but rather to discover more interpretable brain representations and systematically evaluate their explanations. Since individual voxels may respond to overlapping visual concepts, we introduce sparse autoencoders (SAEs) for fMRI decomposition to help disentangle these representations into interpretable patterns. Unlike voxel-level explanations, these patterns capture joint activity across multiple voxels while remaining directly interpretable in terms of their contributing voxels and spatial distribution. (ii) Scalability. BrainExplore separates the computationally expensive hypothesis-generation stage from subsequent interpretation and evaluation. By constructing a reusable, brain-inspired dictionary of approximately 1,300 concepts and precomputing image–concept labels, we can efficiently evaluate hundreds of thousands of candidate components across decomposition methods, ROIs, and subjects, without requiring separate VLM/LLM explanations for each pattern. (iii) Quantitative interpretability and validation. BrainExplore provides substantially more extensive and quantitative validation than previous approaches, comparing decomposition methods, brain regions, and subjects, with validation on held-out measured fMRI, alignment with known ROI functionality, cross-subject consistency, encoder robustness, and LAION-fMRI generalization. We further added cross-dataset concept-consistency experiments (see mJJL-Q4) and following the reviewer's suggestion, contrastive specificity evaluations controlling for correlated semantic and low-level visual properties (see W2).

New comparison with BrainSCUBA and LaVCa. We have conducted an additional comparison on NSD Subject 1 across EBA, FFA, and PPA. We mapped their generated voxel captions to our fixed concept dictionary and evaluated all methods using the same image labels, ranking/evaluation splits, and scoring procedure. On held-out measured fMRI, BrainExplore (SAE+ICA) recovered 239/62 concepts at alignment thresholds of 0.5/0.8, compared with 86/12 for direct voxel search, 51/10 for LaVCa, and 31/1 for BrainSCUBA. While these results show substantially broader concept recovery by BrainExplore, we emphasize that this is an approximate comparison due to differences in the methods' objectives and adaptations to their implementations (see new Appendix J.5).

W2/Q1. Limited validation of concept specificity. [...] Could the authors evaluate explanations using matched positive and negative images to distinguish the target concept from correlated visual features? This would provide stronger evidence of fine-grained neural selectivity.

Response. Following the reviewer's suggestion, we conducted a new contrastive evaluation to test whether discovered patterns remain selective for their associated concepts after controlling for correlated semantic and low-level visual properties. Specifically, we test whether patterns distinguish their associated concepts from other images sharing potentially confounding properties, e.g., distinguishing food from non-food images that both contain plates.

We introduce two groups of contrastive controls: (i) Semantic concepts: We test whether each pattern distinguishes its assigned concept from other strongly associated ones. (ii) Low-level visual properties: We compute 52 image features across five families (color, spatial color, brightness/contrast, edges/orientation, and spatial frequency) and test whether the pattern remains selective when each property is held approximately constant. Both analyses use conditional ROC-AUC on held-out measured and predicted fMRI (see new Appendix J.6). Among concepts with original evaluation scores above 0.8, 70.0% on measured fMRI and 90.6% on predicted fMRI passed both the semantic and all five low-level control families. For concepts with scores above 0.5, the rates were lower, consistent with their weaker initial associations. The lower pass rates on measured fMRI also partly reflect limited coverage of specific concept combinations, reducing statistical power. These results provide further evidence that many discovered representations remain selective for their associated concepts beyond correlated semantic and low-level visual properties.


W3 (minor). The spatial localization of SAE patterns without any explicit spatial constraints seems interesting. However, this can arise from correlations between neighboring voxels due to fMRI measurement or preprocessing, rather than distinct neural representations. Could the authos quantify the localization and compare it against spatial-correlation-preserving controls while accounting for sparsity differences across decomposition methods?

Response. Following the reviewer's suggestion, we quantified the spatial localization of SAE patterns and compared them with ICA. We considered patterns with alignment scores ≥0.8 in three representative ROIs (EBA, FFA-1, and PPA) for Subject 1. For each pattern, we retained the top 10% of ROI vertices ranked by absolute value, ensuring matching counts across methods. As shown in the table below (average across ROIs), SAE patterns exhibit substantially greater spatial clustering than ICA patterns, as reflected by fewer connected components, larger largest-component fractions, and larger mean component sizes.

Method# Components (median) ↓Largest-component fraction (median) ↑Mean component size ↑
SAE28.722.5%9.5
ICA110.85.6%2.5

To assess whether the observed localization could arise from spatial correlations rather than the specific patterns discovered by SAE, we constructed two correlation-based controls using fMRI training data. A smoothness-matched control reproduces local correlations between neighboring voxels, whereas a full-covariance control preserves the broader correlation structure across all pairs of voxels. SAE patterns remained more spatially clustered than both methods. Additional methodological details and results are provided in Appendix J.7.

We thank the reviewer again for the time and effort devoted to reviewing our paper and for their constructive feedback. We hope our additional analyses and responses have addressed your questions and concerns. We would be happy to address any remaining concerns or provide further clarification if needed.

Reviewer_N5j7OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper focuses on uncovering the functional atlas of the human visual cortex. To this end, the authors first conduct analyses using the NSD dataset and synthetic data. Methodologically, for distinct visual ROIs, they adopt unsupervised dimensionality reduction techniques such as sparse autoencoders to learn voxel representations, and interpret the magnitudes of these sparse representations as activation strengths. Next, the authors select 16 images with high activation under a specific activation pattern, and leverage a vision-language model to summarize their shared visual semantics, enabling the identification and interpretation of neural activity patterns. The primary contribution of this work is an automated framework for interpreting fMRI signals in terms of visual semantics. However, its findings rest largely on statistical correlations and semantic scoring by the VLM, and independent neuroscientific experiments are still lacking to verify whether these patterns correspond to genuine, stable functional representations.

Critical strengths
  • The problem investigated in this work — data-driven discovery of human brain functional atlases — is of good value and represents a promising direction for AI for Neuroscience.
Critical weaknesses
  • This study is limited to data-driven observations, and the fine-grained functional atlases reported lack closed-loop validation via real-stimulus experiments. Its validity remains unclear.

  • To interpret a single visual concept, the paper selects only 16 images, which appears insufficient. Since images in NSD are sampled from MSCOCO and consist of complex natural scenes where each image contains multiple intertwined semantics, the semantics distilled from merely 16 images may lack stability.

  • The paper performs unsupervised dimensionality reduction on voxels using sparse autoencoders, which offer limited interpretability. This hinders the establishment of a clear relationship between visual semantics and voxel-level function.

Questions for authors
  • How well does the sparse autoencoder perform on fMRI data? Given its sparsity constraint, could it discard critical information during voxel-wise dimensionality reduction?
  • Some works, such as [1], evaluate encoding performance via image reconstruction based on fMRI features. Could such an evaluation paradigm be adopted here?
  • Are the analytical findings consistent when the SAE is trained on different subsets of data? One possible robustness check is to split the fMRI data used for SAE training into two non-overlapping partitions, train separate SAEs on each partition, and repeat the downstream analyses to verify the stability of the conclusions.

[1] V. S. Wadia et al. A shared code for perceiving and imagining objects in human ventral temporal cortex.

5 comments

We thank the reviewer for their time, effort, and constructive feedback. We conducted additional experiments addressing their suggestions and responded to each point below. Full experimental details and results are provided in the temporary Appendix J for the reviewer's convenience and will be integrated into the relevant sections of the final paper.


W1. This study is limited to data-driven observations, and the fine-grained functional atlases reported lack closed-loop validation via real-stimulus experiments. Its validity remains unclear.

Response. While our approach is indeed data-driven, we would like to emphasize that our findings are extensively evaluated using real, experimentally measured fMRI responses to natural images. Specifically, the discovered pattern–hypothesis associations are evaluated on held-out NSD measurements that were not used for selecting these associations. This enables us to validate discovered representations without requiring new fMRI data collection, which is costly and accessible to relatively few research groups.

Moreover, we demonstrate generalization to held-out images (see new experiment in W2), consistency across subjects (Sec. 4.3), and generalization to the independent LAION-fMRI dataset (Appendix E). We have now additionally evaluated cross-dataset concept consistency (see new Appendix J.1), finding that 73.5% of high-scoring LAION ROI–concept associations were recovered in NSD using measured fMRI, increasing to 94.4% with combined measured and predicted fMRI. Recovery in the reverse direction was lower, consistent with differences in dataset coverage discussed in Appendix E. While definitive validation will require targeted follow-up experiments, we believe our extensive evaluations already provide substantial evidence for the reliability and generalizability of the discovered representations.


W2. To interpret a single visual concept, the paper selects only 16 images, which appears insufficient. Since images in NSD are sampled from MSCOCO and consist of complex natural scenes where each image contains multiple intertwined semantics, the semantics distilled from merely 16 images may lack stability.

Response. We thank the reviewer for raising this important concern. While the original evaluation on measured fMRI relies on a relatively small number of top-activating images, the combined evaluation includes a substantially larger set of predicted responses (approximately 120 top-activating images per pattern). In addition to the held-out generalization experiment in Appendix B.6, we have now conducted a larger-scale evaluation across thousands of images to further assess the reliability of the discovered representations.

In this new experiment, we ask whether a discovered brain pattern consistently responds more strongly to images containing its associated concept than to images without it. To test this, we select the best SAE pattern for each concept using the ranking split, then evaluate the fixed pattern–concept association on all 4,921 measured fMRI responses and 60,000 predicted responses held out from pair selection. We quantify this using ROC-AUC, which measures how well pattern activation distinguishes images containing the concept from those without it, with 0.5 indicating chance. Across 363 concepts with ranking alignment scores above 0.5, the median ROC-AUC was 0.77 on measured fMRI and 0.85 on predicted fMRI. For the 73 concepts with scores above 0.8, these increased to 0.81 and 0.88, respectively. All results were statistically significant against 10,000 label permutations (Holm-adjusted p = 0.00040). These findings demonstrate that the discovered concept–pattern associations generalize to much larger image sets, including real measured brain responses (see more details in new Appendix J.2).

W3. The paper performs unsupervised dimensionality reduction on voxels using sparse autoencoders, which offer limited interpretability. This hinders the establishment of a clear relationship between visual semantics and voxel-level function.

Response. We would like to clarify that our SAE is not used for conventional dimensionality reduction. Rather, we employ an overcomplete latent representation (expansion factor 4), together with sparsity constraints, to help disentangle potentially overlapping visual representations. While individual voxels provide a direct anatomical correspondence, their activity may reflect multiple overlapping visual concepts. Our goal is therefore to discover representations distributed across voxels, rather than assign a single function to each voxel. Importantly, each learned SAE component corresponds to a weighted pattern over the original voxels, allowing us to directly visualize its spatial distribution and identify the contributing voxels. Quantitatively, SAE discovers substantially more interpretable concepts than individual voxels (104 vs. 42 using measured fMRI alone; Table 1), demonstrating the benefit of this decomposition. We further assess how well the SAE preserves the original fMRI information in our response to Q1.


Q1/Q2. How well does the sparse autoencoder perform on fMRI data? Given its sparsity constraint, could it discard critical information during voxel-wise dimensionality reduction? Some works, such as [1] (V. S. Wadia et al. A shared code for perceiving and imagining objects in human ventral temporal cortex.), evaluate encoding performance via image reconstruction based on fMRI features. Could such an evaluation paradigm be adopted here?

Response. As clarified in W3, our SAE uses an overcomplete representation, rather than conventional dimensionality reduction. Nevertheless, sparsity can potentially lead to information loss, and we therefore evaluate how well the SAE preserves the original fMRI information. First, in Appendix B.6, we evaluate fMRI reconstruction on held-out measured responses, achieving a high median Pearson correlation of 0.78. Following the reviewer's suggestion to assess the preservation of visual information, we conducted an image-retrieval experiment to test whether SAE-reconstructed fMRI responses retain sufficient information to identify the viewed image. While image reconstruction requires a separate generative model, retrieval provides a direct quantitative alternative. Given a measured fMRI response and 1,000 candidate images, we first predict the fMRI response for each candidate using a trained image-to-fMRI encoder. We then rank the candidates by their Pearson correlation with the measured response and repeat the ranking using its SAE reconstruction instead. The SAE-reconstructed responses achieved 94.3% Top-5 retrieval accuracy, compared with 97.7% using the original measured responses (see new Appendix J.3). While some information is lost during reconstruction, these results demonstrate that the SAE preserves substantial stimulus-related information, supporting its effectiveness for discovering interpretable visual representations.


Q3. Are the analytical findings consistent when the SAE is trained on different subsets of data? One possible robustness check is to split the fMRI data used for SAE training into two non-overlapping partitions, train separate SAEs on each partition, and repeat the downstream analyses to verify the stability of the conclusions.

Response. Following the reviewer's suggestion, we divided the SAE training data into two non-overlapping subsets of equal size, trained independent SAEs with identical architectures and hyperparameters, and repeated the concept-scoring analysis across all ROIs. The resulting concept scores showed strong agreement (Spearman rank correlation of 0.87). We further tested whether high-scoring concepts (≥0.8) discovered using one SAE were also recovered by the independently trained SAE at a threshold of 0.5. We found near-perfect recovery in both directions: 77/77 (100%) from Split 1 to Split 2 and 61/62 (98.4%) in the reverse direction. These results demonstrate that the discovered concept associations are highly robust to the choice of SAE training subset (see new Appendix J.4).

We thank the reviewer again for their detailed and constructive review. We hope our additional experiments and responses have addressed your questions and concerns. If any concerns remain, we would be happy to address them before the discussion period ends.

Comments on the paper

1 comment

We sincerely thank all reviewers for their time, effort, and constructive engagement throughout the review process. We are pleased that all reviewers have expressed support for accepting our paper. We particularly appreciate their detailed feedback and suggestions, which motivated additional experiments and clarifications that substantially strengthened the paper.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.