Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

MindAligner: Explicit Brain Functional Alignment for Cross-Subject Visual Decoding from Limited fMRI Data

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read MindAligner claims that a low-rank linear map between subjects' fMRI spaces lets a frozen decoder reconstruct a new subject's visual experience from one hour of data.

desk verdict Plausible explicit brain-alignment method with a real training-inference mismatch that needs a direct validation of the transfer matrix before the central claim holds. read the letter →

arxiv 2502.05034 v1 pith:RGTK2FHI submitted 2025-02-07 cs.CV

Yuqin Dai , Zhouheng Yao , Chunfeng Song , Qihao Zheng , Weijian Mai , Kunyu Peng , Shuai Lu , Wanli Ouyang
show 2 more authors
Jian Yang Jiamin Wu
This is my paper · ORCID
classification cs.CV
keywords braindecodingcross-subjectfMRIalignmentvisualreconstructionTransferMatrixfunctionaldata-limitedNaturalScenesDataset
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

MindAligner tries to show that cross-subject visual decoding does not require per-subject retraining of a large generative decoder: a lightweight linear map between brain spaces can carry a new subject's fMRI signal into a known subject's space, where a frozen pre-trained decoder already works. The paper's setting is deliberately hard: only one hour of fMRI from the new subject, and no paired trials in which two subjects viewed the same image. It reports that this explicit alignment beats implicit shared-latent-space baselines on most reconstruction and retrieval metrics, with only about 6% of the full model's parameters trained. If correct, the result makes new-subject brain decoding cheap enough for clinical and brain-computer-interface use, and the learned map doubles as a tool for studying which brain regions vary most across people.

What carries the argument

The core object is the Brain Transfer Matrix (BTM), a low-rank linear map $M = A B$ with hidden dimension $h = 4096$ that projects a novel subject's fMRI voxel space into a known subject's voxel space. Training this matrix is carried by the Brain Functional Alignment Module, whose cross-stimulus neural mapper uses feature-wise linear modulation (FiLM), conditioned on the difference between CLIP embeddings of two similar but non-identical images, to synthesize a known-subject fMRI embedding from a novel-subject embedding. Multi-level alignment then enforces signal-level reconstruction fidelity between synthesized and real fMRI, distributional consistency through KL divergence, and semantic consistency through a latent alignment loss that compares fMRI-embedding dissimilarities with image-embedding dissimilarities. The machinery matters because it produces paired training signal without requiring the two subjects to have viewed the same image, which is exactly the constraint real fMRI datasets impose.

What would settle it

Train the pipeline with the cross-stimulus neural mapper replaced by a fixed identity mapping or by random conditioning vectors; if decoding performance stays at the reported level, the gain does not come from functional alignment. Alternatively, test the trained Brain Transfer Matrix on a held-out block where both subjects viewed the same image and check whether the predicted known-subject fMRI $\hat{F}^K$ matches the true same-stimulus fMRI better than chance.

Watch

Extended reading notes

Core claim

The paper's central claim is that explicit functional alignment in the original voxel space is sufficient for cross-subject visual decoding under data scarcity. It learns a Brain Transfer Matrix $M = A B$ that maps a novel subject's fMRI $F^N$ into a known subject's space as $\hat{F}^K = M F^N$; because the matrix is decomposed into two low-rank linear factors, adapting to a new subject trains only a small fraction of the full model's parameters. Since the dataset lacks paired same-stimulus trials, the matrix is trained with a cross-stimulus neural mapper that modulates the latent fMRI embedding by the CLIP embedding difference between two similar images, supervised by a reconstruction loss, a KL divergence loss, and a latent alignment loss based on stimulus similarity. At inference only the transfer matrix is kept, and the projected signal feeds the frozen pre-trained decoder. The paper reports that this outperforms implicit-alignment baselines on most decoding and retrieval metrics, and that the transfer weights reveal graded cross-subject variability, with early visual cortex relatively conserved and higher visual and ventral-stream regions varying most.

Load-bearing premise

The load-bearing premise is that the cross-stimulus neural mapper can synthesize a known subject's fMRI from a novel subject's fMRI under a different but similar image, conditioned only on the CLIP embedding difference between the two images; if that synthesis is unfaithful, the Brain Transfer Matrix is trained on fabricated pairs and the reported gains could come from the conditioning signal rather than from genuine brain alignment.

Editorial extensions

If this is right

  • A new subject needs only about one hour of fMRI: the pre-trained decoder stays frozen, and only the transfer matrix plus training-only auxiliary modules are learned, so adaptation costs roughly 6% of the original trainable parameters.
  • Explicit one-to-one voxel-space alignment outperforms implicit shared-latent alignment: on the paper's single-session setting it reports higher fMRI spatial correlation and better retrieval, including a 17.9% improvement in brain retrieval over the baseline.
  • The choice of which known subject to align to has little effect on reconstructed images, which the paper interprets as evidence that the alignment captures shared visual function rather than a particular subject's idiosyncrasies.
  • The transfer weights provide a region-level map of cross-subject variability: early visual cortex is more conserved, while higher visual cortex and the ventral pathway regions such as PPA, FFA, OPA, and EBA show larger inter-subject differences.
  • Alignment quality remains similar under data scarcity: the transfer-quantity distribution computed from one hour of data resembles the distribution from the full dataset, so the interpretability claims hold in the data-limited regime.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper, if the transfer matrix truly encodes stable functional correspondences, it should be possible to build a population-level transfer map from a few fully scanned subjects and apply it to a completely new subject without any of that subject's own fMRI; the paper only demonstrates the one-hour-data version.
  • Beyond the paper, the claimed variability hierarchy, with early visual cortex conserved and the ventral stream more variable, is a testable neuroscience prediction: in any dataset where two subjects view identical images, direct inter-subject voxel correlations should show the same ordering, otherwise the pattern may reflect the stimulus-conditional mapper rather than true brain organization.
  • Beyond the paper, a linear voxel-to-voxel map is modality-agnostic once trained; the same transfer matrix could in principle be plugged into other frozen fMRI decoders, for example for video, imagined speech, or memory, a step the paper does not take but its architecture permits.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes MindAligner, a method for cross-subject fMRI visual decoding in a data-limited setting (one hour of new-subject data). It learns a low-rank Brain Transfer Matrix M = A × B that maps a novel subject's voxel responses into a known subject's voxel space, allowing a pre-trained frozen decoder to reconstruct visual stimuli. Because the training sessions lack strictly paired identical stimuli, a Brain Functional Alignment module uses a cross-stimulus neural mapper (FiLM conditioned on CLIP embedding differences between similar images) to synthesize pseudo-pairs, with signal-level (L_rec, L_KL) and latent (L_latent) alignment losses. At inference, the mapper is discarded and only M is used. Experiments on NSD compare with MindEye2 and MindBridge, reporting consistent gains especially in brain retrieval (75.3% vs 57.4%), ablations, efficiency (about 6% trainable parameters), and brain-heatmap analyses of inter-subject variability.

Significance. If the central claim holds, MindAligner is a valuable contribution: explicit, lightweight, voxel-level cross-subject alignment that does not require retraining a large per-subject decoder and yields interpretable functional correspondences. The paper's strengths include the low-rank factorization, a training scheme that avoids shared-stimulus requirements, a public code commitment, and falsifiable neuroscience-oriented analyses (TQ and fSC visualizations) that are consistent with known visual hierarchy. However, the current evidence does not yet fully establish the central claim: the separation between the conditioned training path and the unconditional inference path is not validated, and the quantitative comparisons lack statistical support. These issues are addressable with additional experiments, but they are load-bearing for the paper's main message.

major comments (3)
  1. [§4.3 (Eqs. 3–6) vs §4.4 (Eq. 1)] The Brain Transfer Matrix is trained only along the conditioned path. In Section 4.3, the synthesized known-subject signal is computed as F_K_hat = MC(z_N, E_diff) × B, with z_N = A × F_N, so every gradient that updates A and B flows through the cross-stimulus neural mapper. Section 4.4 then discards the mapper and applies the unconditional map F_K_hat = (A × B) F_N, a path that no loss term directly supervises. Consequently, nothing in the optimization forces A × B alone to be a valid cross-subject map, and the reported inference gains could be attributable to image-specific information encoded in the conditioning signal E_diff during training. This is the load-bearing gap for the central claim. Please add an ablation that trains the BTM without the mapper, i.e., minimizes L_rec, L_KL, and L_latent directly on the unconditional output (A × B) F_N, or otherwise provide direct evidence (e.g., reconstruction loss on a validation set) that the unconditional path is accurate.
  2. [Tables 1 and 7] No error bars, confidence intervals, or significance tests are reported for any decoding metric. The per-subject rows in Table 7 show that MindAligner is not uniformly better than MindEye2 (e.g., subject 1: SSIM 0.415 vs 0.428 and Image↑ 90.90% vs 93.96%; subject 7: SSIM 0.407 vs 0.408), so the text's claim that MindAligner 'outperforms existing methods' or 'surpasses the baseline in almost all metrics' is stronger than the point estimates warrant. Please report standard errors or bootstrap intervals across subjects and sessions, and run paired significance tests on at least the headline metrics (e.g., PixCorr, SSIM, Image↑, Brain↑). Qualify the 'almost all metrics' statements accordingly.
  3. [Eq. 9] The latent alignment loss is confounded by the conditioning variable. The FiLM input is E_diff = E_N − E_K (Eq. 3), and the target in Eq. 9 is the dissimilarity R(E_N, E_K) between the same two CLIP embeddings. Since z_K = MC(z_N, E_diff), the mapper can reduce L_latent by encoding E_diff into z_K without learning a stable, subject-generic functional correspondence; the same confounding applies to the full objective because no loss evaluates the unconditional path. Please add a control experiment in which the conditioning input is shuffled, randomized, or held fixed across pairs, and report whether L_latent and the final decoding performance are preserved. This is needed to support the interpretation that the BTM, not the conditioning signal, is responsible for the reported brain alignment.
minor comments (5)
  1. [Eq. 8] The KL divergence between two high-dimensional voxel vectors is not defined as written; specify how F_K_hat and F_K are converted to probability distributions (e.g., softmax over voxels or over batches).
  2. [§5.3] The fSC metric description is too brief: 'Pearson correlation between corresponding brain regions' leaves unclear which regions are used, how correspondence is established, and whether the correlation is computed on raw or aligned signals.
  3. [§2.2] Thual et al. (2023) and Ferrante et al. (2024) are cited as explicit alignment methods, but neither is included in the quantitative comparison; please state whether their evaluation protocols are compatible with the one-hour setting and, if so, add them to Table 1.
  4. [Appendix B, Table 5] The table caption contains the typo 'trainalble'; also, the appendix would benefit from stating the hidden size h = 4096 in the main text next to the parameter counts.
  5. [§5.5] The notation '1 → 2' is introduced in the caption of Figure 6 but used before a formal definition; define it in the main text of Section 5.5.

Circularity Check

2 steps flagged · score 4.0 of 10

The held-out BTM-alone decoding result is genuine, but the L_latent alignment loss (Eq. 9) is self-definitional — its target is a function of the same CLIP-embedding difference fed to the mapper as its condition — and the BTM is only ever trained and ablated through that conditioned FiLM path, so the paper's 'explicit alignment' attribution is partially circular.

  1. self definitional [Section 4.3, Eqs. (3)-(5) and Eq. (9)]
    "we leverage the differences between I N and I K as conditions for generating fMRI... Ediff = Eimage(IN ) − Eimage(IK), (3)... zK = MC(zN , zdiff), (5)... we design a latent alignment loss Llatent by enforcing the consistency between fMRI embedding pairs and stimuli pairs: Llatent = ||(R(Ef (zN ), Ef (zK)) − R(EN , EK)||2 2, (9) where EN and EK denote the CLIP embeddings of IN and IK."

    The L_latent target R(EN, EK) is a dissimilarity between exactly the two CLIP embeddings whose difference is fed to the cross-stimulus mapper as its conditioning signal (E_diff = E_image(I_N) − E_image(I_K), z_diff = E_diff × M_diff, z_K = MC(z_N, z_diff)). The target is thus a function of the model's own conditioning input: a FiLM layer can drive this loss toward zero by injecting E_diff into the generated embedding (e.g., Ef(z_K) ≈ Ef(z_N) + E_diff W) without learning any stable cross-subject functional correspondence in A×B.

  2. other [Sections 4.2-4.4, Eqs. (1), (6), (10); Table 2]
    "To learn a reliable brain transfer matrix, the Brain Functional Alignment Module (BFA) conducts soft cross-subject alignment in both the brain space and the shared latent space of the BTM... The transformed embedding zK is then projected to the known subject's space by the low-rank matrix B... ˆF K = zK × B. (6)... the final brain alignment loss LAlign is formulated as: LAlign = LDec+αrecLrec+αKL LKL +αlaLlatent, (10)... During inference, only the trained BTM is used for functional alignment (Eq. 1)."

    Every term of Eq. (10) is evaluated on the conditioned output of Eq. (6), F̂_K = MC(A F_N, E_diff M_diff) B, whose FiLM condition carries the image difference E_diff; none of the four losses is computed on the unconditional operator deployed at inference, F̂_K = A × B × F_N (Eq. 1), and the Table 2 ablation never removes MC. The paper nonetheless concludes from this objective that it learns 'a reliable brain transfer matrix' and attributes the decoding gains to 'the effective learning of brain transfer matrix that accurately aligns the novel subject.' The stated equations train only the composite A∘MC(·,E_diff)∘B; whether M alone performs the alignment is a skipped link, not a consequence of the objective.

full rationale

MindAligner's headline result — cross-subject decoding with the BTM alone (M = A×B, Eq. 1) from one hour of data, outperforming MindEye2 and MindBridge — is evaluated with the deployed model on held-out shared images against external baselines, so the core claim does not reduce to its inputs. L_rec (Eq. 7) targets real known-subject fMRI, the fSC and TQ analyses are direct measurements of the trained matrix, and the two self-citations (Mai et al., 2023; Mai et al., 2024) appear only in related-work enumerations and are not load-bearing. Two issues nonetheless weaken the 'explicit functional alignment' derivation. First, L_latent (Eq. 9) is self-definitional: its target R(E_N, E_K) is a function of the very CLIP embeddings whose difference (Eq. 3) is fed to the cross-stimulus mapper as its conditioning input, so injecting E_diff into the generated embedding minimizes the loss without encoding any cross-subject correspondence in A×B. Second, every loss in Eq. (10) evaluates the MC-conditioned path (Eq. 6) while inference deploys the unconditional path (Eq. 1); no loss or ablation supervises the deployed matrix, so the attribution of decoding gains to the BTM itself is asserted rather than derived, and the conditioning signal can absorb image-specific information during training. Because the final decoding numbers are honest measurements with M alone, the central claim retains independent empirical content; the circularity is component-level (the latent-alignment loss) plus a training/deployment mismatch, not a collapse of the headline result. Score 4.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the linear transfer assumption, the validity of synthetic cross-stimulus pairs, and the use of CLIP geometry as a brain-similarity proxy. These are domain assumptions rather than derived facts, and none is independently verified in the paper. The main free parameter beyond standard training choices is the BTM rank h, with loss weights as additional hand-set coefficients.

free parameters (2)
  • BTM hidden size h = 4096
    Rank of the low-rank decomposition M = A×B (Eq. 2); chosen by hand. Appendix D shows performance drops at h=64 and is comparable at h=1024, so this capacity choice affects results.
  • Loss weights alpha_rec, alpha_la, alpha_KL, alpha_1, alpha_2 = 1, 0.001, 0.001, 0.033, 0.016
    Coefficients in Eq. 10; set by hand and not learned. The ablation in Table 2 shows the loss combination matters.
assumptions (5)
  • domain assumption There exists a low-rank linear map M = A × B from novel subject voxel space to known subject voxel space that preserves decoding-relevant information.
    Eqs. 1 and 2; the entire BTM approach assumes linear transferability of fMRI responses at voxel level.
  • ad hoc to paper fMRI responses under similar but non-identical visual stimuli can be transformed into each other by a linear modulation conditioned on CLIP embedding differences.
    Sec. 4.3, Eqs. 3-6; this cross-stimulus synthesis is specific to MindAligner and is not validated independently.
  • domain assumption CLIP embedding dissimilarity is a valid proxy for cross-subject brain functional dissimilarity.
    Eq. 9; L_latent uses CLIP semantics as the target geometry for aligning fMRI embeddings.
  • domain assumption A pre-trained decoding model for a known subject can decode aligned novel-subject fMRI without retraining.
    Sec. 4.4; the inference procedure feeds the aligned signal directly into the frozen decoder.
  • domain assumption The one-session sample is representative enough to estimate the transfer matrix.
    Sec. 5.2; the data-limited claim depends on a single hour of fMRI being sufficient to fit M.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MindAligner: Explicit Brain Functional Alignment for Cross-Subject Visual Decoding from Limited fMRI Data." pith.science (2026). https://pith.science/paper/RGTK2FHI

@misc{pith2026250205034,
  author       = {Pith},
  title        = {Pith review of: MindAligner: Explicit Brain Functional Alignment for Cross-Subject Visual Decoding from Limited fMRI Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RGTK2FHI}},
  note         = {Machine review of arXiv:2502.05034}
}
read the original abstract

Brain decoding aims to reconstruct visual perception of human subject from fMRI signals, which is crucial for understanding brain's perception mechanisms. Existing methods are confined to the single-subject paradigm due to substantial brain variability, which leads to weak generalization across individuals and incurs high training costs, exacerbated by limited availability of fMRI data. To address these challenges, we propose MindAligner, an explicit functional alignment framework for cross-subject brain decoding from limited fMRI data. The proposed MindAligner enjoys several merits. First, we learn a Brain Transfer Matrix (BTM) that projects the brain signals of an arbitrary new subject to one of the known subjects, enabling seamless use of pre-trained decoding models. Second, to facilitate reliable BTM learning, a Brain Functional Alignment module is proposed to perform soft cross-subject brain alignment under different visual stimuli with a multi-level brain alignment loss, uncovering fine-grained functional correspondences with high interpretability. Experiments indicate that MindAligner not only outperforms existing methods in visual decoding under data-limited conditions, but also provides valuable neuroscience insights in cross-subject functional analysis. The code will be made publicly available.

Figures

Figures reproduced from arXiv: 2502.05034 by the authors.

Figure 1
Figure 1. Different approaches to functional alignment in brain decoding: Prior works (Scotti et al., 2024b; Li et al., 2025) adopt implicit alignment approach that aligns all subjects into a single latent space, which may lead to suboptimal alignment. Differently, MindAligner employs an explicit alignment strategy, mapping novel subject signals to seen ones by establishing fine-grained functional correspondences. MindAligner… view at source ↗
Figure 2
Figure 2. Overview of MindAligner. To achieve explicit brain functional alignment, given a pre-trained brain decoding model, we design a Brain Functional Alignment Module (BFA) that learns a Brain Transfer Matrix (BTM) M for fMRI mapping between the known and novel subjects. BTM is decomposed into two low-rank matrices A and B to create latent space for further alignment. The Cross-Stimulus Neural Mapper is proposed to create… view at source ↗
Figure 3
Figure 3. Visualization of MindAligner’s decoding results from training on one hour of data. Stimulus 5→1 5→2 5→7 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Visualization of transfer quantity in brain heatmaps [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Comparison of fSC results between MindAligner and the baseline. detailed cross-subject results in Appendix C. Computational Efficiency. We compare the computational efficiency between our model and baseline MindEye2 w.r.t. parameter count and inference time per image. …
Figure 7
Figure 7. Figure 7: Visualization of transfer quantity in brain heatmaps from MindAligner training using 1 hour and full datasets. functional alignment strategy with the baseline in reducing inter-subject differences, as shown in [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: More visualization of brain decoding results under different novel and known subjects [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Fast Whole-Brain, Geometry-Aware Functional Alignment for Cross-Subject Decoding

    q-bio.NC 2026-07 conditional novelty 6.5 of 10

    SpectralOT regularizes entropic optimal transport with the first three Laplace-Beltrami eigenmodes of cortical geometry to produce fast, parsimonious whole-brain functional alignments that improve cross-subject decoding.

  2. CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding

    cs.HC 2025-06 conditional novelty 6.0 of 10

    CSBrain, a cross-scale spatiotemporal EEG foundation model, reports state-of-the-art results across 11 tasks and 16 datasets, but its claim of consistent superiority over all baselines is not supported by its own tables.

  3. Spiking Neural Networks for fMRI-Based Visual Semantic Decoding

    cs.NE 2026-07 conditional novelty 5.0 of 10

    Spiking-neural-network image features are more predictable from fMRI activity under a fixed linear decoder than ResNet-18 features, but the effect is inflated by feature-scale differences.

Reference graph

Works this paper leans on

48 extracted references · 28 canonical work pages · cited by 3 Pith papers

  1. [1]

    J., St-Yves, G., Wu, Y., Breedlove, J

    Allen, E. J., St-Yves, G., Wu, Y., Breedlove, J. L., Prince, J. S., Dowdle, L. T., Nau, M., Caron, B., Pestilli, F., Charest, I., et al. A massive 7t fmri dataset to bridge cognitive neuroscience and artificial intelligence. Nature neuroscience, 25 0 (1): 0 116--126, 2022

    work page 2022

  2. [2]

    Wills aligner: A robust multi-subject brain representation learner

    Bao, G., Gong, Z., Zhang, Q., Zhou, J., Fan, W., Yi, K., Naseem, U., Hu, L., and Miao, D. Wills aligner: A robust multi-subject brain representation learner. arXiv preprint arXiv:2404.13282, 2024

    arXiv 2024

  3. [3]

    Local optimal transport for functional brain template estimation

    Bazeille, T., Richard, H., Janati, H., and Thirion, B. Local optimal transport for functional brain template estimation. 2019. CorpusID:162169103

    work page 2019

  4. [4]

    An empirical evaluation of functional alignment using inter-subject decoding

    Bazeille, T., DuPre, E., Richard, H., Poline, J., and Thirion, B. An empirical evaluation of functional alignment using inter-subject decoding. NeuroImage, 245: 0 118683, 2021. doi:10.1016/j.neuroimage.2021.118683. URL https://www.sciencedirect.com/science/article/pii/S1053811921009563

    work page arXiv 2021

  5. [5]

    Unsupervised learning of visual features by contrasting cluster assignments

    Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., and Joulin, A. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33: 0 9912--9924, 2020

    2020

  6. [6]

    Seeing beyond the brain: Conditional diffusion model with sparse masked modeling for vision decoding

    Chen, Z., Qing, J., Xiang, T., Yue, W., and Zhou, J. Seeing beyond the brain: Conditional diffusion model with sparse masked modeling for vision decoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 22710--22720, 2023

    work page 2023

  7. [7]

    Conroy, B., Singer, B., Haxby, J., and Ramadge, P. J. fmri-based inter-subject cortical alignment using functional connectivity. Advances in neural information processing systems, 22, 2009

    work page 2009

  8. [8]

    J., Wassermann, D., Thirion, B., and Mensch, A

    Dadi, K., Varoquaux, G., Machlouzarides-Shalit, A., Gorgolewski, K. J., Wassermann, D., Thirion, B., and Mensch, A. Fine-grain atlases of functional modes for fmri analysis. NeuroImage, 221: 0 117126, 2020

    work page 2020

Show all 48 references
  1. [9]

    Through their eyes: multi-subject brain decoding with simple alignment techniques

    Ferrante, M., Boccato, T., Ozcelik, F., VanRullen, R., and Toschi, N. Through their eyes: multi-subject brain decoding with simple alignment techniques. Imaging Neuroscience, 2: 0 1--21, 2024

    2024

  2. [10]

    S., Scheinost, D., Finn, D

    Finn, E. S., Scheinost, D., Finn, D. M., Shen, X., Papademetris, X., and Constable, R. T. Can brain state be manipulated to emphasize individual differences in functional connectivity? NeuroImage, 160: 0 140--151, 2017

    2017

  3. [11]

    Mind-3d: Reconstruct high-quality 3d objects in human brain

    Gao, J., Fu, Y., Wang, Y., Qian, X., Feng, J., and Fu, Y. Mind-3d: Reconstruct high-quality 3d objects in human brain. arXiv preprint arXiv:2312.07485, 2023

    2023 arXiv

  4. [12]

    fmri-3d: A comprehensive dataset for enhancing fmri-based 3d reconstruction

    Gao, J., Fu, Y., Wang, Y., Qian, X., Feng, J., and Fu, Y. fmri-3d: A comprehensive dataset for enhancing fmri-based 3d reconstruction. arXiv preprint arXiv:2409.11315, 2024

    2024 arXiv

  5. [13]

    Mindtuner: Cross-subject visual decoding with visual fingerprint and semantic correction

    Gong, Z., Zhang, Q., Bao, G., Zhu, L., Liu, K., Hu, L., and Miao, D. Mindtuner: Cross-subject visual decoding with visual fingerprint and semantic correction. arXiv preprint arXiv:2404.12630, 2024

    2024 arXiv

  6. [14]

    Generative adversarial networks

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. Generative adversarial networks. Communications of the ACM, 63 0 (11): 0 139--144, 2020

    2020

  7. [15]

    Decoding natural image stimuli from fmri data with a surface-based convolutional network

    Gu, Z., Jamison, K., Kuceyeski, A., and Sabuncu, M. Decoding natural image stimuli from fmri data with a surface-based convolutional network. arXiv preprint arXiv:2212.02409, 2022

    2022 arXiv

  8. [16]

    Denoising diffusion probabilistic models

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 0 6840--6851, 2020

    2020

  9. [17]

    and Kamitani, Y

    Horikawa, T. and Kamitani, Y. Generic decoding of seen and imagined objects using hierarchical visual features. Nature communications, 8 0 (1): 0 15037, 2017

    2017

  10. [18]

    Mindshot: Brain decoding framework using only one image

    Jiang, S., Meng, Z., Liu, D., Li, H., Su, F., and Zhao, Z. Mindshot: Brain decoding framework using only one image. arXiv preprint arXiv:2405.15278, 2024

    2024

  11. [19]

    Krizhevsky, A., Sutskever, I., and Hinton, G. E. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012

    2012

  12. [20]

    Unibrain: Universal brain mri diagnosis with hierarchical knowledge-enhanced pre-training

    Lei, J., Dai, L., Jiang, H., Wu, C., Zhang, X., Zhang, Y., Yao, J., Xie, W., Zhang, Y., Li, Y., et al. Unibrain: Universal brain mri diagnosis with hierarchical knowledge-enhanced pre-training. arXiv preprint arXiv:2309.06828, 2023

    2023 arXiv

  13. [21]

    Enhancing cross-subject fmri-to-video decoding with global-local functional alignment

    Li, C., Qian, X., Wang, Y., Huo, J., Xue, X., Fu, Y., and Feng, J. Enhancing cross-subject fmri-to-video decoding with global-local functional alignment. In European Conference on Computer Vision, pp.\ 353--369. Springer, 2025

    2025

  14. [22]

    Lin, S., Sprague, T., and Singh, A. K. Mind reader: Reconstructing complex images from brain activities. Advances in Neural Information Processing Systems, 35: 0 29624--29636, 2022

    2022

  15. [23]

    Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Doll \'a r, P., and Zitnick, C. L. Microsoft coco: Common objects in context. In Computer Vision--ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pp....

    2014

  16. [24]

    Minddiffuser: Controlled image reconstruction from human brain activity with semantic and structural diffusion

    Lu, Y., Du, C., Zhou, Q., Wang, D., and He, H. Minddiffuser: Controlled image reconstruction from human brain activity with semantic and structural diffusion. In Proceedings of the 31st ACM International Conference on Multimedia, pp.\ 5899--5908, 2023

    2023

  17. [25]

    and Zhang, Z

    Mai, W. and Zhang, Z. Unibrain: Unify image reconstruction and captioning all in one diffusion model from human brain activity. arXiv preprint arXiv:2308.07428, 2023

    2023 arXiv

  18. [26]

    Brain-conditional multimodal synthesis: A survey and taxonomy

    Mai, W., Zhang, J., Fang, P., and Zhang, Z. Brain-conditional multimodal synthesis: A survey and taxonomy. IEEE Transactions on Artificial Intelligence, 2024

    2024

  19. [27]

    N., Nishimoto, S., and Gallant, J

    Naselaris, T., Kay, K. N., Nishimoto, S., and Gallant, J. L. Encoding and decoding in fmri. Neuroimage, 56 0 (2): 0 400--410, 2011

    2011

  20. [28]

    and VanRullen, R

    Ozcelik, F. and VanRullen, R. Brain-diffuser: Natural scene reconstruction from fmri signals using generative latent diffusion. arxiv 2023. arXiv preprint arXiv:2303.05334

    2023 arXiv

  21. [29]

    and VanRullen, R

    Ozcelik, F. and VanRullen, R. Natural scene reconstruction from fmri signals using generative latent diffusion. Scientific Reports, 13 0 (1): 0 15666, 2023

    2023

  22. [30]

    Reconstruction of perceived images from fmri patterns and semantic brain exploration using instance-conditioned gans

    Ozcelik, F., Choksi, B., Mozafari, M., Reddy, L., and VanRullen, R. Reconstruction of perceived images from fmri patterns and semantic brain exploration using instance-conditioned gans. In 2022 International Joint Conference on Neural Networks (IJCNN), pp.\ 1--8. IEEE, 2022

    2022

  23. [31]

    Film: Visual reasoning with a general conditioning layer

    Perez, E., Strub, F., De Vries, H., Dumoulin, V., and Courville, A. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018

    2018

  24. [32]

    Binless kernel machine: Modeling spike train transformation for cognitive neural prostheses

    Qian, C., Sun, X., Wang, Y., Zheng, X., Wang, Y., and Pan, G. Binless kernel machine: Modeling spike train transformation for cognitive neural prostheses. Neural Computation, 32 0 (10): 0 1863--1900, 2020

    1900

  25. [33]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In ICML, pp.\ 8748--8763. PMLR, 2021

    2021

  26. [34]

    Hierarchical text-conditional image generation with clip latents

    Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., and Chen, M. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 0 (2): 0 3, 2022

    2022 arXiv

  27. [35]

    Brain decoding of the human connectome project tasks in a dense individual fmri dataset

    Rastegarnia, S., St-Laurent, M., DuPre, E., Pinsard, B., and Bellec, P. Brain decoding of the human connectome project tasks in a dense individual fmri dataset. NeuroImage, 283: 0 120395, 2023

    2023

  28. [36]

    Reconstructing the mind's eye: fmri-to-image with contrastive learning and diffusion priors

    Scotti, P., Banerjee, A., Goode, J., Shabalin, S., Nguyen, A., Dempster, A., Verlinde, N., Yundler, E., Weisberg, D., Norman, K., et al. Reconstructing the mind's eye: fmri-to-image with contrastive learning and diffusion priors. Advances in Neural Information Processing Syste...

    2024

  29. [37]

    S., Tripathy, M., Villanueva, C

    Scotti, P. S., Tripathy, M., Villanueva, C. K. T., Kneeland, R., Chen, T., Narang, A., Santhirasegaran, C., Xu, J., Naselaris, T., and Norman, K. A. e. a. Mindeye2: Shared-subject models enable fmri-to-image with 1 hour of data. In ICML, 2024 b

    2024

  30. [38]

    Generative adversarial networks for reconstructing natural images from brain activity

    Seeliger, K., Güçlü, U., Ambrogioni, L., Güçlütürk, Y., and Van Gerven, M. Generative adversarial networks for reconstructing natural images from brain activity. NeuroImage, 181: 0 775--785, 2018

    2018

  31. [39]

    Neuro-vision to language: Enhancing brain recording-based visual reconstruction and language interaction

    Shen, G., Zhao, D., He, X., Feng, L., Dong, Y., Wang, J., Zhang, Q., and Zeng, Y. Neuro-vision to language: Enhancing brain recording-based visual reconstruction and language interaction. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

  32. [40]

    Deep image reconstruction from human brain activity

    Shen, G., Horikawa, T., Majima, K., and Kamitani, Y. Deep image reconstruction from human brain activity. PLoS computational biology, 15 0 (1): 0 e1006633, 2019

    2019

  33. [41]

    Rethinking the inception architecture for computer vision

    Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., and Wojna, Z. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 2818--2826, 2016

    2016

  34. [42]

    and Nishimoto, S

    Takagi, Y. and Nishimoto, S. High-resolution image reconstruction with latent diffusion models from human brain activity. pp.\ 14453--14463, 2023

    2023

  35. [43]

    and Le, Q

    Tan, M. and Le, Q. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, pp.\ 6105--6114. PMLR, 2019

    2019

  36. [44]

    H., Zemskova, T., Courty, N., Flamary, R., Dehaene, S., and Thirion, B

    Thual, A., Tran, Q. H., Zemskova, T., Courty, N., Flamary, R., Dehaene, S., and Thirion, B. Aligning individual brains with fused unbalanced gromov wasserstein. Advances in neural information processing systems, 35: 0 21792--21804, 2022

    2022

  37. [45]

    Aligning brain functions boosts the decoding of visual semantics in novel subjects

    Thual, A., Benchetrit, Y., Geilert, F., Rapin, J., Makarov, I., Banville, H., and King, J.-R. Aligning brain functions boosts the decoding of visual semantics in novel subjects. arXiv preprint arXiv:2312.06467, 2023

    2023 arXiv

  38. [46]

    Mindbridge: A cross-subject brain decoding framework

    Wang, S., Liu, S., Tan, Z., and Wang, X. Mindbridge: A cross-subject brain decoding framework. pp.\ 11333--11342, 2024

    2024

  39. [47]

    C., Sheikh, H

    Wang, Z., Bovik, A. C., Sheikh, H. R., and Simoncelli, E. P. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13 0 (4): 0 600--612, 2004

    2004

  40. [48]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.