Brain Diffusion for Visual Exploration: Cortical Discovery using Large Scale Generative Models

A long standing goal in neuroscience has been to elucidate the functional organization of the brain. Within higher visual cortex, functional accounts have remained relatively coarse, focusing on regions of interest (ROIs) and taking the form of selectivity for broad categories such as faces, places, bodies, food, or words. Because the identification of such ROIs has typically relied on manually assembled stimulus sets consisting of isolated objects in non-ecological contexts, exploring functional organization without robust a priori hypotheses has been challenging. To overcome these limitations, we introduce a data-driven approach in which we synthesize images predicted to activate a given brain region using paired natural images and fMRI recordings, bypassing the need for category-specific stimuli. Our approach -- Brain Diffusion for Visual Exploration ("BrainDiVE") -- builds on recent generative methods by combining large-scale diffusion models with brain-guided image synthesis. Validating our method, we demonstrate the ability to synthesize preferred images with appropriate semantic specificity for well-characterized category-selective ROIs. We then show that BrainDiVE can characterize differences between ROIs selective for the same high-level category. Finally we identify novel functional subdivisions within these ROIs, validated with behavioral data. These results advance our understanding of the fine-grained functional organization of human visual cortex, and provide well-specified constraints for further examination of cortical organization using hypothesis-driven methods.

Paper

References (78)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer Vg6A6/10 · confidence 4/52023-07-04

Summary

This paper proposes a diffusion-based method for fMRI-guided image synthesis and claims that it can identify the selectivity of various ROIs in the visual cortex. The motivation of this work is well-defined and is of vital importance to the computational neuroscience field. While the method in this work is intuitive and lacks machine learning depth.

Strengths

Strengths: * Interesting scientific field and question of neuroscience. * The paper is well-written and has a smooth and concrete flow. * The experiments and visualizations are in good condition and sufficient.

Weaknesses

Weaknesses * The method used in this paper is heuristic and lacks technical depth. * Generating or reconstructing images from fMRI is not a new idea, as has been proposed in [1], etc. * CLIP is an off-the-shelf large-scale pre-trained model, performing conditional diffusion-based generation with embedding from CLIP or Stable Diffusion is stereotyped [2,3]. * In Section 3.3, the domain-specific techniques you designed for this specific task are just some first-order tricks (i.e., linear combination of S, euler approximation). Maybe these tricks are effective enough, but that's overly empirical. * The experimental section is empirical and lacks theoretical analysis. Your analysis of the fMRI encoder with ROIs is interesting, but little focus had been put on what's the scientific relationship of neuroscience and diffusion model. [1] https://www.biorxiv.org/content/10.1101/2022.11.18.517004v3 [2] https://arxiv.org/abs/2112.10752 [3] https://arxiv.org/pdf/2208.01618.pdf

Questions

Please consider the things listed in the “Weaknesses” section. Also please consider providing information regarding the interpretability of the utilization of the diffusion model in your context.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

3 good

Limitations

To my knowledge and understanding, there is no potential negative societal impact of this work. Other limitations please see “Weaknesses”.

Reviewer SV7e7/10 · confidence 5/52023-07-04

Summary

The paper presents a new algorithm to guide a diffusion model to decode maximally activating images for particular voxel subregions of human fMRI using an encoding model trained to predict brain activity from images. The algorithm identifies stereotypic features for defined ROIs, such as faces or food items. The authors evaluate the specificity of the reconstructed image with CLIP zero-shot classification and human evaluation. They also cluster the encoding model weights and find that the clusters result in perceivable semantic categories in the reconstructed images. This algorithm allows for more fine-grained analysis of preferences across the visual system.

Strengths

There have been many papers that reconstruct images from brain activity. However, all of them require some form of retraining the diffusion model. This approach only needs an encoding model that can be trained independently. The extensive human evaluation is great. While this goes beyond what I would ask for a NeurIPS paper, I think it would be great to also verify them by showing them back to human subjects.

Weaknesses

The authors show that the reconstructed images possess stable properties that can be identified by humans and by a CLIP network. However, they don't show that the images also do what they are supposed to do, which is activate a particular subnetwork. While human experiments probably go beyond the scope of this NeurIPS paper, one way to do that would be to take **another** encoding model for the same neural activity, take it as a proxy for the brain and show the images to that.

Questions

Q1: As far as I understand $\gamma$ trades off between the denoising gradient and the maximization gradient. I couldn't find what value you set it too, or how you chose it. Q2: There is a preprint that describes pretty much the same idea as in the paper but uses it for V4 single neurons: Pawel A. Pierzchlewicz, Konstantin F. Willeke, Arne F. Nix, Pavithra Elumalai, Kelli Restivo, Tori Shinn, Cate Nealley, Gabrielle Rodriguez, Saumil Patel, Katrin Franke, Andreas S. Tolias, Fabian H. Sinz Energy Guided Diffusion for Generating Neurally Exciting Images https://www.biorxiv.org/content/10.1101/2023.05.18.541176v1 If one looks at the date, the preprint has been submitted around the NeurIPS deadline, so it's unlikely that the authors were aware of it and I would consider this as a case where two groups had the same idea. However, I think it would be fair to mention/cite them. Q3: Can you discuss the compute time required for a single image?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Limitations are discussed.

Reviewer fpHd7/10 · confidence 4/52023-07-05

Summary

This study proposes BrainDIVE, a system for synthesizing optimal stimuli for any given region of interest in the brain. The model combines a pretrained latent diffusion model for image generation with a linear “brain encoder” trained to map CLIP feature vectors onto the corresponding brain activity. At test time, a gradient-based optimization iteratively produces an image that maximizes the activity of a predefined set of voxels. The technique is validated on well-known ROIs, producing the expected images. It can also highlight subtle differences between ROIs selective to the same broad category, or functional subdivisions of existing ROIs. The functional hypotheses are validated by human subjects evaluating the properties of the generated pictures. The technique appears simple but can provide meaningful information for neuroscience studies.

Strengths

* The method can retrieve subtle differences between ROIs selective for the same class (e.g. OFA and FFA). * The method can highlight functional differences between sub-clusters of existing ROIS (e.g. food clusters) * The qualitative observations are validated by behavioral evaluations from human subjects.

Weaknesses

* The Related works section tends to overstate the novelty of the technique. It only mentions NeuroGen as prior work, whereas other prior studies had also attempted to generate optimal stimuli for specific ROIs: for instance, Ratan Murty et al (2021), Ozcelik et al (2022), Ozcelik et al (2023). The latter also used a diffusion model for image generation, as done here. Furthermore, you seem to be aware of at least some of these studies, since you criticized one of them in your Methods section for using a “hand-derived” prior. It would be much better to properly acknowledge all these studies in the Related works section, and clarify your method’s advantages/disadvantages (e.g. no hand-derived prior/time-consuming iterative method).

Questions

* I do not understand why you systematically need to report the top-5 or top-10 out of 500 generated images? Do I understand correctly that the pool of 500 are all generated for the same objective (ROI maximization)? If yes, then what does this imply about the remaining 99% of images? Are they worse than NeuroGen? Are they actually failure cases? What do the worse 1% of images look like? If they are not representative of the expected category, does that invalidate your optimization method? * Appendix, section 9, training objective: this section is useful to understand the image optimization process. One thing I do not understand from this section or from Figure 2 is why the optimization is performed in the diffusion latent space, rather than in CLIP space? Why do the gradients have to flow all the way into the diffusion model? Couldn’t you iteratively optimize the $CLIP_{img}$ vector for the same objective, and then use it to condition the diffusion model in one pass (as you would with a text prompt)? This should be computationally more efficient, and I don’t see any reason why it would be less accurate.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Limitations are properly acknowledged.

Reviewer MC3H7/10 · confidence 4/52023-07-21

Summary

This paper introduces Brain Diffusion for Visual Exploration (BrainDiVE) that aimed at exploring the fine-grained functional organization of the human visual cortex. Motivated by the limitations of previous studies that relied on researcher-crafted stimuli, BrainDiVE leverages generative deep learning models trained on large-scale image datasets and brain activation data from fMRI recordings. The proposed method uses brain maps as guidance to synthesize diverse and realistic images, enabling data-driven exploration of semantic preferences across visual cortical regions. By applying BrainDiVE to category-selective voxels and individual ROIs, the authors demonstrate its ability to capture semantic selectivity and identify subtle differences in response properties within specific brain networks. It is also shown that BrainDiVE identifies novel functional subdivisions within existing ROIs, highlighting its potential for providing new insights into the human visual system's functional properties.

Strengths

The paper's methodology shows promise in capturing semantic selectivity and identifying fine-grained functional distinctions within visual cortical regions. Also, the paper's potential significance lies in applying BrainDiVE to understand the fine-grained functional organisation of the human visual cortex. By providing insights into category selectivity, response properties, and sub-regional divisions, the paper opens avenues for further exploratory neuroscience studies. To achieve this, the authors perform the experiments in the manuscript are extensive, covering: 1. the semantic specificity of the method by decoding images from task fMRI and literature-obtained ROIs 2. compared the abstraction of face representation in the brain by comparing images decoded from two regions, the fusiform face area (FFA) and the occipital face area (OFA). 3. Use the method to extend knowledge of brain function by finding subdivisions in known areas in the cases of food decoding. The authors describe the experimental setting clearly at the beginning of section 4 for all cases. The experiments show that the manuscript has a good balance between the combination of known methodological approaches and addressing interesting questions in neuroscience. Specifically, the authors perform a commendable effort in obtaining quantitative results using human evaluators (cf Tables 3 and 4). Finally, the authors show through results evidence that their method is a window into understanding region-specificity hypotheses of brain regions which are knowingly involved in different aspects of visual processing.

Weaknesses

The paper could benefit from comparing its results with previous contributions [e.g. 9 and 10] to highlight its novelty and contributions in relation to existing methods. The methodology part lacks clarity, particularly in explaining key components like the diffusion model architecture and brain-guided synthesis—further clarifying the role of the image-to-brain encoder in influencing the denoising process during inference. The generalisability and reliability of results are hard to asses through a small dataset, specifically just 10 subjects. Furthermore, the paper lacks clarity on how the sub-divisions of the visual cortex are being verified or validated. Whether the results are specific to the analysed regions or the algorithm is being too biased by the experimental condition is not clear through the experiments. For instance, what would happen if the authors try to decode an area not specific to visual processing? An example of this would be using subsections of the orbitofrontal gyrus or other brain areas not expected to perform well in reconstructing visual stimuli. As a small point authors should review the presentation of images. Figure 1 shows mostly best-case scenarios which are then not as good in Figure 4 (for instance the case of images generated from face voxels). This might bias readers. Second, some details, such as figure 4 appearing before figure 3 make reading the paper confusing.

Questions

1. Validation of BrainDiVE's Effectiveness: a. Can the authors provide more details on the validation process to ensure that BrainDiVE-generated images indeed effectively activate the targeted brain regions? b. Could the authors consider conducting more extensive comparisons with other existing methods [e.g. 9 and 10], to demonstrate the advantages and uniqueness of BrainDiVE in eliciting specific brain activations? c. Could the authors show the results on decoding a baseline region which is not involved in visual processing? 2. Clarity in Methodology: The methodology part could benefit from more clarity and detailed explanations of key components, such as the diffusion model architecture and the exact implementation of brain-guided synthesis. Improving the architecture in Figure-2 might help. 3. Statistical Analysis of Qualitative Evaluation: As the paper relies on qualitative evaluation with 10 subjects, could the authors mention this explicitly and what limitations are expected from this in the limitations section? 4. Ethical Considerations: Could the authors describe the demographics of the population, or acknowledge the lack of this information to inform of wether the results are biased to a specific gender or population group. 5. Reproducibility: Are the authors intending to release the code and not publicly available data such as the scores produced by the human raters upon acceptance of the paper? This is key to guarantee reproducibility

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Authors have not addressed the negative societal impact of their work, but this can be fixed by adding specific text in the Discussion section. The authors don't mention the demographic characteristics of the small human subject database that they have used not mention any ethical concerns related to decoding images from brain activations. This is fixable so authors should address it.

Authorsrebuttal2023-08-14

We appreciate the request for code. The anonymous github for the paper is now in a publicly visible comment. Also reproduced here for convenience: https://anonymous.4open.science/r/BrainDiVE_code-FF0C

Authorsrebuttal2023-08-09

Link to code

Here is a copy to a fully anonymized version of the core code used in this paper: https://anonymous.4open.science/r/BrainDiVE_code-FF0C The code includes the following: * fMRI encoder architecture * fMRI encoder training * fMRI encoder validation * Diffusion architechture * Diffusion inference using broad category selective regions; ROIs; and parts of ROIs We are fully committed to releasing a complete version of the code at a later point in time.

Reviewer SV7e2023-08-11

Thanks for the clarifications

Dear authors, thank you for your clarifications. I have read the rebuttal and will stand by my evaluation.

Authorsrebuttal2023-08-11

We appreciate the timely response. Thank you again for the positive evaluation of this paper!

Authorsrebuttal2023-08-16

We are deeply grateful for the positive assessment you've given to our paper! Thank you again for your suggestions and comments.

Reviewer MC3H2023-08-19

Thanks for the responses. I have updated my score to “Accept” following the suggestions for the evaluation of a good contribution with impact in a specific field

Authorsrebuttal2023-08-19

We are very grateful for the positive assessment you've given to our work! We would like to again express our appreciation for your suggestions.

Reviewer Vg6A2023-08-19

I thank the authors for a detailed response and for answering my concerns one by one. **Q1.1)** : It's true that the problem that you investigate is different from the Takagi et al. [1]. However, the method and model framework you use is much similar to them. Yours and their methods are both in the Latent Diffusion Model scope. As you mentioned in Section 3.1. **Q1.2)**: Thanks for your comment. This part makes sense to me. Your new technique of performing fMRI-conditioned image synthesis is interesting. **Interpretability**: I think the results of Section 4 of the supplementary is a characteristic of a normal diffusion model, in which it tends to generate the main features of an image in the first steps and then details in latter steps. I wonder how the gradient guidance of your method contributes to this generation process. In conclusion, while there's room for refinement in the methodology, the paper offers significant scientific contribution to neuroscience. I have revised my score in light of these observations. [1] High-resolution image reconstruction with latent diffusion models from human brain activity

Authorsrebuttal2023-08-20

We appreciate the positive evaluation you have given to this paper!

Program Chairsdecision2023-09-21

Decision

Accept (oral)

© 2026 NYSGPT2525 LLC