Pith. sign in

REVIEW 4 major objections 6 minor 49 references

This paper contends that spiking-neural-network features are more recoverable from fMRI activity than conventional ANN features, and that choosing the right target features is itself a key part of visual decoding.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · deepseek-v4-flash

2026-08-01 13:13 UTC pith:BZ5CAO2I

load-bearing objection The 0.77-to-0.03 MSE headline is largely a scale artifact from using one ridge penalty across feature spaces; the paper still has a real, if mixed, empirical story that deserves refereeing. the 4 major comments →

arxiv 2607.19170 v1 pith:BZ5CAO2I submitted 2026-07-21 cs.NE

Spiking Neural Networks for fMRI-Based Visual Semantic Decoding

classification cs.NE
keywords fMRI visual decodingspiking neural networkssemantic decodingvisual featuresridge regressionneural representationimage retrievalbrain-machine interface
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper asks whether the visual features that fMRI decoding tries to predict should be spike-based rather than standard ANN activations. It keeps the same linear decoder and preprocessing fixed, and changes only the target features: dense ResNet-18 features versus temporally averaged spike features from four spiking variants of the same residual backbone. On two public fMRI datasets, the authors report that SNN-derived features are predicted with far lower error (MSE from 0.77 to 0.03 on GoD) and support higher semantic agreement (Top-1 from 0.18 to 0.44). The paper concludes that SNN representations are more brain-decodable targets and that target feature design matters as much as the decoder itself.

Core claim

The authors establish, under a controlled comparison, that spiking-neural-network-derived visual features are more linearly recoverable from fMRI responses and preserve more semantic information than ANN features from the same architectural family. Four SNN variants—LIF, PSN, MPSN, and BuSNN—all outperform the ResNet-18 ANN baseline on feature-prediction error, and most improve semantic classification, retrieval, and semantic-guided reconstruction. The discovery is empirical: the advantage appears when the fMRI-to-feature decoder is held fixed, isolating the choice of target representation as the cause.

What carries the argument

The controlled comparison rests on a fixed L2-regularized linear decoder (ridge regression) that maps fMRI voxel activity to 512-dimensional visual features. The only thing varied across conditions is the prediction target: dense continuous activations from a ResNet-18 ANN, versus spike-rate features from spiking variants of the same residual backbone, averaged across T simulation time steps. The spiking dynamics—membrane integration, parallel computation, memory propagation, or burst coding—generate temporally structured, sparse representations whose decodability is the paper's central object.

Load-bearing premise

The headline comparison assumes that using one fixed L2 penalty (α=1000) is fair for target features with very different numerical scales—SNN spike-rate features near zero versus larger ANN activations—so the reported MSE advantage is not a scale artifact.

What would settle it

Re-run the decoding on GoD and NSD with each target feature space normalized (e.g., z-scored per dimension or scaled to unit variance) before fitting the same ridge decoder; if SNN targets no longer beat ANN targets on MSE, MFC, or Top-1 agreement, the central claim collapses.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • On the GoD dataset, switching to SNN targets lowers feature-prediction MSE from 0.77 to 0.03 and lifts top-1 semantic agreement from 0.18 to 0.44.
  • On NSD, SNN targets also reduce MSE and raise top-1 agreement, with the best variant improving agreement from 0.23 to 0.42.
  • Retrieval improves on GoD: the best SNN variant lifts Acc@1 from 0.44 to 0.58 and Acc@5 from 0.82 to 0.94.
  • Ablations show both spiking neural dynamics and the number of simulation time steps matter; T=4 outperforms T=2 and T=8.
  • Target feature choice becomes a scientific variable in fMRI decoding rather than a fixed engineering default.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • An untested confound is the single L2 penalty applied to target features with very different numerical scales; normalizing each feature space before fitting the same decoder could shrink or reverse the reported MSE gap.
  • The voxel-level alignment advantage suggests SNN features could also improve neural encoding models, a neighboring question the paper does not pursue.
  • A testable extension is to vary regularization strength or feature normalization per target and check whether the ANN-versus-SNN ranking is stable across decoders.
  • If the effect holds, it suggests a design heuristic for future decoding targets: sparse, temporally averaged spike-rate features are more linearly recoverable from BOLD signals than dense activations.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes that SNN-derived visual features, rather than conventional ANN features, should be used as regression targets in fMRI-based visual semantic decoding. Using a fixed ridge-regression decoder and four SNN variants sharing a ResNet-18-style backbone, the authors compare feature-prediction error, semantic top-1 agreement, retrieval accuracy, and semantic-guided reconstruction against a ResNet-18 ANN baseline across GoD, NSD, and Mini-Algonauts 2021. The central claim is that SNN features are more linearly recoverable from fMRI and preserve more semantic information, with headline results including a GoD MSE drop from 0.7707 to 0.0282 and a Top-1 agreement increase from 0.1800 to 0.4400. The paper also includes ablations on spiking neuron dynamics and simulation time steps.

Significance. If the claims were fully validated, the paper would make a useful contribution by reframing target-feature choice as a scientific design variable in fMRI decoding, rather than a fixed engineering default. The controlled comparison across multiple public benchmarks, the use of several spiking dynamics in the same architecture family, and the inclusion of a scale-free metric (MFC) are strengths. However, the headline evidence is not currently scale-invariant, the scale-free metric gives mixed results across SNN variants, and the semantic evaluation is partly self-referential. These issues are load-bearing because they affect the abstract's main quantitative claims. The underlying hypothesis is plausible and worth testing, but the evidence as presented is not yet convincing enough for acceptance.

major comments (4)
  1. [§IV-C, Table I] The headline MSE comparison is not scale-invariant and is confounded by the fixed regularization coefficient α=1000. The targets are not standardized: ResNet-18 penultimate activations are dense and substantially larger in magnitude, while SNN features are temporal averages of binary spikes (Eq. 6), typically in [0,1]. For ridge regression, the same L2 penalty shrinks the ANN mapping much more aggressively relative to its signal, inflating ANN MSE, whereas low-variance SNN targets naturally yield low MSE. The train/test pattern in Table I—ANN train MSE 0.139 vs test 0.771 (5.5× degradation), while SNN train/test are both below 0.07—is consistent with this scale artifact. The paper's claim that the comparison is fair because the decoder and regularization are fixed (last paragraph of §III-C) overlooks that ridge regularization interacts with target scale. Please re-run the comparison with
  2. [abstract and §IV-C, Table I] The scale-free MFC metric does not support the broad claim that SNN-derived features exhibit stronger alignment with fMRI responses or are more readily recoverable. On GoD, MPSN has MFC 0.2583, below ANN's 0.2624; on NSD, LIF has 0.1506 and MPSN has 0.1929, both below ANN's 0.2064. Only PSN and BuSNN consistently exceed ANN on this metric. The text itself acknowledges that 'LIF and MPSN show lower MFC despite their lower MSE,' but the abstract and conclusion still assert a general SNN advantage. Please either qualify the central claim to the specific SNN variants and metrics that support it, or provide additional scale-free evidence (e.g., per-dimension correlation, rank-correlation, or a normalized-target MSE) that shows a consistent advantage.
  3. [§III-D, Table II] The Top-1 'semantic decoding accuracy' is not accuracy against ground-truth object labels; it is agreement between the classifier prediction on the fMRI-decoded feature and the classifier prediction on the model-derived feature of the same backbone. Because the target features and the classifier head are from the same model, a target space that is easier to regress onto, or that produces more confident predictions from the model's own head, will mechanically increase this score. This makes the 0.1800→0.4400 improvement difficult to interpret as an improvement in semantic decoding of the actual viewed category. Please also report Top-1 accuracy against the true image categories, and clearly distinguish it from agreement in the text and abstract. In addition, Table II contains an unexplained anomaly: on NSD, LIF and BuSNN have train Top-1 agreement of 0.10000, while their validation and te
  4. [§IV-E, Table III and §IV-F, Table IV] Retrieval and reconstruction results are reported as single point estimates without confidence intervals or significance tests. On a small test set, differences such as Acc@1 0.44 vs 0.58 (Table III) or Bo5-SRC 0.7402 vs 0.7553 (Table IV) may be within sampling noise. Please provide bootstrap confidence intervals or permutation tests for Acc@K and Bo5-SRC, and report the number of test samples. Also, the Bo5-SRC metric (Eq. 11) uses hand-assigned weights (0.5, 0.3, 0.2); please add a sensitivity analysis or report the three components separately, since the overall score can be dominated by arbitrary weight choices.
minor comments (6)
  1. [§II, references] Reference [33] and [40] are the same paper (Horikawa and Kamitani, Nature Communications 2017); [13] and [52] are also duplicated. Please consolidate.
  2. [Fig. 2] The voxel-level alignment uses one-tailed Welch's t-tests across many comparisons (5 ROIs × 4 SNN variants), but no multiple-comparison correction is reported. Please state the number of subjects/samples and whether the significance levels survive correction.
  3. [§IV-C, Eq. (10)] For MFC, dimensions with standard deviation <1e-8 are excluded, but the number of excluded dimensions is not reported. If a large fraction of SNN feature dimensions are near-constant, the metric may be computed on only a small subset. Please report how many dimensions are valid for each feature space.
  4. [§IV-G, Fig. 5] The time-step ablation varies T over {2,4,8} but only for LIF. The choice of T=4 as the default is justified only retrospectively; consider a small validation-based selection or a sensitivity discussion.
  5. [§IV-B] The paper says '10% of the training samples from each dataset as a validation set for model selection' but the decoder's only hyperparameter α is fixed. Please clarify what quantity is used for model selection, or whether the validation set is only used for early stopping/evaluation.
  6. [§IV-F, Eq. (11)] The definition of C_i as 'normalized cosine similarity in feature space' is ambiguous: normalized relative to what? Please define it so that the score is reproducible.

Circularity Check

1 steps flagged

Top-1 'semantic decoding accuracy' is defined as agreement with the same model that supplies the regression targets, so the headline semantic gain reduces to the feature-reconstruction gain; the rest of the comparison is largely self-contained.

specific steps
  1. self definitional [Section III-D (Semantic Classification, Retrieval, and Reconstruction) and Table II; abstract 'top-1 semantic decoding accuracy']
    "For semantic decoding, z is passed to the classifier head of the corresponding visual backbone. ... The same classifier head is also applied to the model-derived feature of the stimulus, so that semantic predictions from fMRI-decoded and model-derived features can be compared. ... Top-1 agreement measures whether the most confident category predicted from the fMRI-decoded feature matches that predicted from the corresponding model-derived feature."

    The 'ground truth' for the headline semantic-decoding number is not the stimulus category but the prediction of the very backbone that also defines the regression targets z^(m) in Eq. (8). If the fMRI decoder were perfect, z = z^(m), then passing both through the same fixed classifier head would force Top-1 agreement to 1 by construction for any backbone. Therefore Table II is essentially a label-level restatement of the feature-prediction MSE in Table I: a target space that is easier to regress will automatically show higher 'semantic decoding accuracy.' The abstract's 0.1800 -> 0.4400 claim is thus the feature-reconstruction improvement renamed, not an independent semantic benchmark.

full rationale

The paper's main comparison is an empirical, controlled decoder study on external fMRI datasets (GoD, NSD, Mini-Algonauts), with fixed preprocessing, fixed decoder form, and frozen ImageNet backbones. The core regression result (Eq. 7-9) and the voxel-level PCC analysis are self-contained statistical procedures, not derivations from first principles, so no circularity arises there. The fixed ridge penalty α=1000 across target spaces of different numerical scale is a serious validity confound for the MSE comparison, but it is a measurement- or calibration artifact, not a circularity: the SNN MSE advantage is not logically forced by the paper's equations. Self-citations are present (e.g., [25], [31] for MPSN and BuSNN, [23], [24] with a co-author), but they supply candidate neuron models and SNN-training machinery rather than loading the paper's central argument, so they are not load-bearing circularity. The one genuinely circular element is the 'semantic decoding accuracy' metric: it compares fMRI-decoded features with the same model's own classification of the same image, so the evaluation oracle and the regression target are identical by construction. Retrieval on GoD (Table III) and reconstruction (Table IV) provide some independent external grounding, which is why the overall circularity score is moderate rather than severe.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The paper relies on standard pretrained encoders and public fMRI datasets. Its main free parameters are the shared ridge penalty (α=1000), the SNN simulation time step (T=4), and an ad hoc reconstruction-metric weighting. The fixed-α choice is the most consequential because it makes the MSE comparison scale-sensitive.

free parameters (3)
  • ridge regularization coefficient α = 1000
    Ridge regularization strength fixed across all feature spaces; because target feature scales differ, this hand-set parameter confounds the ANN–SNN comparison.
  • SNN simulation time steps T = 4
    Default number of time steps set to 4; ablation shows T=2 and T=8 perform worse, so T=4 is effectively a tuned choice.
  • Bo5-SRC metric weights = 0.5 Ci + 0.3 Ri + 0.2 Ui
    Weights in Eq (11) for the reconstruction score are defined ad hoc; no justification or sensitivity analysis is given.
axioms (4)
  • domain assumption Linear ridge-regression decodability is a sufficient probe for comparing feature spaces.
    The entire comparison rests on treating linear predictability from fMRI as the measure of decodability (Sec III-C); a nonlinear decoder could change the ranking.
  • ad hoc to paper The same regularization coefficient is appropriate for differently scaled target features.
    Fixed α=1000 across ANN/SNN targets has no stated justification; this is load-bearing for the MSE comparison (Sec III-C).
  • domain assumption SNN features averaged over time steps are comparable in information content to a single ANN activation vector.
    Temporal averaging Eq (6) is used to create fixed-length targets; no check that temporal averaging preserves semantic info equally across variants.
  • domain assumption Pretrained ImageNet encoders provide suitable semantic feature spaces for fMRI decoding.
    Follows prior work, but the conclusion about 'brain-decodable' representations is mediated by these model spaces.

pith-pipeline@v1.3.0-alltime-deepseek · 14359 in / 13961 out tokens · 150329 ms · 2026-08-01T13:13:57.156336+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Spiking Neural Networks for fMRI-Based Visual Semantic Decoding." pith.science (2026). https://pith.science/paper/BZ5CAO2I

@misc{pith2026260719170,
  author       = {Pith},
  title        = {Pith review of: Spiking Neural Networks for fMRI-Based Visual Semantic Decoding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BZ5CAO2I}},
  note         = {Machine review of arXiv:2607.19170}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Functional magnetic resonance imaging (fMRI)-based visual decoding aims to recover visual information from measured brain activity, commonly by mapping fMRI responses into latent visual features for downstream decoding tasks. Most existing methods learn mappings from fMRI responses to visual features extracted by artificial neural networks (ANNs), yet it remains unclear whether ANN-derived features provide suitable targets for brain decoding. In this study, we investigate spiking neural network (SNN)-derived visual features as alternative targets for fMRI-based visual decoding. We compare an ANN baseline with four SNN variants from the same architectural family, which differ in their spiking dynamics. To isolate the effect of the target features, all models use the same L2-regularized linear fMRI-to-feature decoder, while only the feature vectors used as regression targets are varied. Compared with the ANN baseline, SNN-derived features exhibit stronger alignment with fMRI responses and improve visual semantic decoding performance. For instance, on the GoD dataset, SNN-derived features reduce feature-prediction error from 0.7707 to 0.0282 and improve top-1 semantic decoding accuracy from 0.1800 to 0.4400. Ablation results further indicate that both spiking neural dynamics and temporal simulation steps contribute to the observed advantage. These findings support SNN-derived features as effective brain-decodable visual representations and highlight target feature design as an important component of fMRI-based visual decoding.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 5 linked inside Pith

  1. [1]

    What we can do and what we cannot do with fmri,

    N. K. Logothetis, “What we can do and what we cannot do with fmri,”Nature, vol. 453, no. 7197, pp. 869–878, 2008

  2. [2]

    Identifying natural images from human brain activity,

    K. N. Kay, T. Naselaris, R. J. Prenger, and J. L. Gallant, “Identifying natural images from human brain activity,”Nature, vol. 452, no. 7185, pp. 352–355, 2008

  3. [3]

    Reconstructing visual experiences from brain activity evoked by natural movies,

    S. Nishimoto, A. T. Vu, T. Naselaris, Y . Benjamini, B. Yu, and J. L. Gallant, “Reconstructing visual experiences from brain activity evoked by natural movies,”Current Biology, vol. 21, no. 19, pp. 1641–1646, 2011

  4. [4]

    Decoding of visual infor- mation from human brain activity: A review of fmri and eeg studies,

    R. Zafar, A. S. Malik, N. Kamel, S. C. Dass, J. M. Abdullah, F. Reza, and A. H. Abdul Karim, “Decoding of visual infor- mation from human brain activity: A review of fmri and eeg studies,”Journal of Integrative Neuroscience, vol. 14, no. 02, pp. 155–168, 2015

  5. [5]

    Cogformer: A unified multi- scale brain representation for visual decoding and reconstruction from fmri,

    X. Yin, J. Q. Gan, and H. Wang, “Cogformer: A unified multi- scale brain representation for visual decoding and reconstruction from fmri,”IEEE Transactions on Medical Imaging, 2026

  6. [6]

    Bayesian reconstruction of natural images from human brain activity,

    T. Naselaris, R. J. Prenger, K. N. Kay, M. Oliver, and J. L. Gallant, “Bayesian reconstruction of natural images from human brain activity,”Neuron, vol. 63, no. 6, pp. 902–915, 2009

  7. [8]

    Alleviating the semantic gap for generalized fmri-to-image reconstruction,

    T. Fang, Q. Zheng, and G. Pan, “Alleviating the semantic gap for generalized fmri-to-image reconstruction,”Advances in Neural Information Processing Systems, vol. 36, pp. 15 096–15 107, 2023

  8. [9]

    Re- trieving and reconstructing conceptually similar images from fmri with latent diffusion models and a neuro-inspired brain decoding model,

    M. Ferrante, T. Boccato, L. Passamonti, and N. Toschi, “Re- trieving and reconstructing conceptually similar images from fmri with latent diffusion models and a neuro-inspired brain decoding model,”Journal of Neural Engineering, vol. 21, no. 4, p. 046001, 2024

  9. [10]

    Deep image reconstruction from human brain activity,

    G. Shen, T. Horikawa, K. Majima, and Y . Kamitani, “Deep image reconstruction from human brain activity,”PLoS Com- putational Biology, vol. 15, no. 1, p. e1006633, 2019

  10. [11]

    Generative adversarial networks for reconstruct- ing natural images from brain activity,

    K. Seeliger, U. G ¨uc ¸l¨u, L. Ambrogioni, Y . G¨uc ¸l¨ut¨urk, and M. A. Van Gerven, “Generative adversarial networks for reconstruct- ing natural images from brain activity,”NeuroImage, vol. 181, pp. 775–785, 2018

  11. [12]

    High-resolution image reconstruc- tion with latent diffusion models from human brain activity,

    Y . Takagi and S. Nishimoto, “High-resolution image reconstruc- tion with latent diffusion models from human brain activity,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 14 453–14 463

  12. [14]

    Neural encoding for human visual cortex with deep neural networks learning “what

    H. Wang, L. Huang, C. Du, D. Li, B. Wang, and H. He, “Neural encoding for human visual cortex with deep neural networks learning “what” and “where”,”IEEE Transactions on Cognitive and Developmental Systems, vol. 13, no. 4, pp. 827–840, 2020

  13. [15]

    Exploring the brain-like properties of deep neural networks: a neural encoding perspective,

    Q. Zhou, C. Du, and H. He, “Exploring the brain-like properties of deep neural networks: a neural encoding perspective,”Ma- chine Intelligence Research, vol. 19, no. 5, pp. 439–455, 2022

  14. [16]

    Beyond brain decoding: Visual-semantic reconstructions to mental creation extension based on fmri,

    H. Jing, D. Jiang, Y . Ma, H. Hua, B. Huang, and N. Zheng, “Beyond brain decoding: Visual-semantic reconstructions to mental creation extension based on fmri,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 19 258–19 268

  15. [17]

    Mindaligner: Explicit brain functional alignment for cross-subject visual decoding from limited fmri data,

    Y . Dai, Z. Yao, C. Song, Q. Zheng, W. Mai, K. Peng, S. Lu, W. Ouyang, J. Yang, and J. Wu, “Mindaligner: Explicit brain functional alignment for cross-subject visual decoding from limited fmri data,”arXiv preprint arXiv:2502.05034, 2025

  16. [18]

    Exploring the visual feature space for multimodal neural decoding,

    W. Xia and C. Oztireli, “Exploring the visual feature space for multimodal neural decoding,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 4370– 4379. 10 IEEE TRANSACTIONS ON MEDICAL IMAGING

  17. [19]

    Seeing through the brain: New insights from decoding visual stimuli with fmri,

    Z. Huang, E. Zhang, W. Qiu, Y . Cai, C. Yang, E. Chen, X. Zhang, R. Ying, D. Zhou, and Y . Yan, “Seeing through the brain: New insights from decoding visual stimuli with fmri,” arXiv preprint arXiv:2510.16196, 2025

  18. [20]

    Bridging brains and concepts: Interpretable visual decoding from fmri with semantic bottlenecks,

    S. Cammarota, M. Ferrante, and N. Toschi, “Bridging brains and concepts: Interpretable visual decoding from fmri with semantic bottlenecks,”Advances in Neural Information Processing Sys- tems, vol. 38, pp. 86 820–86 850, 2026

  19. [21]

    Networks of spiking neurons: the third generation of neural network models,

    W. Maass, “Networks of spiking neurons: the third generation of neural network models,”Neural Networks, vol. 10, pp. 1659– 1671, 1997

  20. [22]

    Deep learning in spiking neural networks,

    A. Tavanaei, M. Ghodrati, S. R. Kheradpisheh, T. Masquelier, and A. Maida, “Deep learning in spiking neural networks,” Neural Networks, vol. 111, pp. 47–63, 2019

  21. [23]

    Spatio-temporal backpropagation for training high-performance spiking neural networks,

    Y . Wu, L. Deng, G. Li, J. Zhu, and L. Shi, “Spatio-temporal backpropagation for training high-performance spiking neural networks,”Frontiers in Neuroscience, vol. 12, p. 331, 2018

  22. [24]

    Direct training for spiking neural networks: Faster, larger, better,

    Y . Wu, L. Deng, G. Li, J. Zhu, Y . Xie, and L. Shi, “Direct training for spiking neural networks: Faster, larger, better,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 33, 2019, pp. 1311–1318

  23. [25]

    Enhancing robustness of spiking neural networks through retina-like coding and memory-based neurons,

    J. Zhang, K. Wang, M. Yao, H. Xu, P. Zhou, B. Xu, and G. Li, “Enhancing robustness of spiking neural networks through retina-like coding and memory-based neurons,”Neural Net- works, p. 107950, 2025

  24. [26]

    Deep residual learning in spiking neural networks,

    W. Fang, Z. Yu, Y . Chen, T. Huang, T. Masquelier, and Y . Tian, “Deep residual learning in spiking neural networks,”Advances in Neural Information Processing Systems, vol. 34, pp. 21 056– 21 069, 2021

  25. [27]

    Stage- wise robust distillation for spiking neural network training,

    S. Shen, J. Zhang, Z. Chen, X. Xia, Y . Hu, and W. Fu, “Stage- wise robust distillation for spiking neural network training,” in ICASSP 2026-2026 IEEE International Conference on Acous- tics, Speech and Signal Processing, 2026, pp. 2421–2425

  26. [28]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, 2016, pp. 770–778

  27. [29]

    Gerstner, W

    W. Gerstner, W. M. Kistler, R. Naud, and L. Paninski,Neuronal dynamics: From single neurons to networks and models of cognition. Cambridge University Press, 2014

  28. [30]

    Parallel spiking neurons with high efficiency and ability to learn long-term dependencies,

    W. Fang, Z. Yu, Z. Zhou, D. Chen, Y . Chen, Z. Ma, T. Masque- lier, and Y . Tian, “Parallel spiking neurons with high efficiency and ability to learn long-term dependencies,”Advances in Neural Information Processing Systems, vol. 36, pp. 53 674– 53 687, 2023

  29. [31]

    Burst spiking neural networks,

    J. Zhang, S. Shen, M. Yao, H. Xu, M. Huang, Y . Tian, B. Xu, and G. Li, “Burst spiking neural networks,”arXiv preprint arXiv:2607.11914, 2026

  30. [32]

    The algonauts project 2021 challenge: How the human brain makes sense of a world in motion,

    R. M. Cichy, K. Dwivedi, B. Lahner, A. Lascelles, P. Iamshchin- ina, M. Graumann, A. Andonian, N. Murty, K. Kay, G. Roig et al., “The algonauts project 2021 challenge: How the human brain makes sense of a world in motion,”arXiv preprint arXiv:2104.13714, 2021

  31. [34]

    A massive 7t fmri dataset to bridge cognitive neuroscience and artificial intelligence,

    E. J. Allen, G. St-Yves, Y . Wu, J. L. Breedlove, J. S. Prince, L. T. Dowdle, M. Nau, B. Caron, F. Pestilli, I. Charestet al., “A massive 7t fmri dataset to bridge cognitive neuroscience and artificial intelligence,”Nature Neuroscience, vol. 25, no. 1, pp. 116–126, 2022

  32. [35]

    Surrogate gradient learning in spiking neural networks,

    E. O. Neftci, H. Mostafa, and F. Zenke, “Surrogate gradient learning in spiking neural networks,”IEEE Signal Processing Magazine, vol. 36, no. 6, pp. 51–63, 2019

  33. [36]

    Training spiking neural networks using lessons from deep learning,

    J. K. Eshraghian, M. Ward, E. O. Neftci, X. Wang, G. Lenz, G. Dwivedi, M. Bennamoun, D. S. Jeong, and W. D. Lu, “Training spiking neural networks using lessons from deep learning,”Proceedings of the IEEE, vol. 111, no. 9, pp. 1016– 1054, 2023

  34. [37]

    Towards spike-based ma- chine intelligence with neuromorphic computing,

    K. Roy, A. Jaiswal, and P. Panda, “Towards spike-based ma- chine intelligence with neuromorphic computing,”Nature, vol. 575, no. 7784, pp. 607–617, 2019

  35. [38]

    Distributed and overlapping representations of faces and objects in ventral temporal cortex,

    J. V . Haxby, M. I. Gobbini, M. L. Furey, A. Ishai, J. L. Schouten, and P. Pietrini, “Distributed and overlapping representations of faces and objects in ventral temporal cortex,”Science, vol. 293, no. 5539, pp. 2425–2430, 2001

  36. [39]

    Visual image reconstruction from human brain activity using a combination of multiscale local image decoders,

    Y . Miyawaki, H. Uchida, O. Yamashita, M.-a. Sato, Y . Morito, H. C. Tanabe, N. Sadato, and Y . Kamitani, “Visual image reconstruction from human brain activity using a combination of multiscale local image decoders,”Neuron, vol. 60, no. 5, pp. 915–929, 2008

  37. [40]

    Generic decoding of seen and imagined objects using hierarchical visual features,

    T. Horikawa and Y . Kamitani, “Generic decoding of seen and imagined objects using hierarchical visual features,”Nature Communications, vol. 8, no. 1, p. 15037, 2017

  38. [41]

    Decoding visual neural representations by multimodal learning of brain-visual-linguistic features,

    C. Du, K. Fu, J. Li, and H. He, “Decoding visual neural representations by multimodal learning of brain-visual-linguistic features,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 9, pp. 10 760–10 777, 2023

  39. [42]

    Brainclip: Bridging brain and visual-linguistic representation via clip for generic natural visual stimulus decoding,

    Y . Liu, Y . Ma, W. Zhou, G. Zhu, and N. Zheng, “Brainclip: Bridging brain and visual-linguistic representation via clip for generic natural visual stimulus decoding,”arXiv preprint arXiv:2302.12971, 2023

  40. [43]

    Natural scene reconstruction from fmri signals using generative latent diffusion,

    F. Ozcelik and R. VanRullen, “Natural scene reconstruction from fmri signals using generative latent diffusion,”Scientific Reports, vol. 13, no. 1, p. 15666, 2023

  41. [44]

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

    P. Scotti, A. Banerjee, J. Goode, S. Shabalin, A. Nguyen, A. Dempster, N. Verlinde, E. Yundler, D. Weisberg, K. Norman et al., “Reconstructing the mind’s eye: fmri-to-image with contrastive learning and diffusion priors,”Advances in Neural Information Processing Systems, vol. 36, pp. 24 705–24 728, 2023

  42. [45]

    Dream: Visual decoding from reversing human visual system,

    W. Xia, R. De Charette, C. Oztireli, and J.-H. Xue, “Dream: Visual decoding from reversing human visual system,” inPro- ceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 8226–8235

  43. [46]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and F.-F. Li, “Imagenet: A large-scale hierarchical image database,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2009, pp. 248–255

  44. [47]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,”Advances in Neural Information Processing Systems, vol. 33, pp. 6840–6851, 2020

  45. [48]

    Spikingjelly: An open- source machine learning infrastructure platform for spike-based intelligence,

    W. Fang, Y . Chen, J. Ding, Z. Yu, T. Masquelier, D. Chen, L. Huang, H. Zhou, G. Li, and Y . Tian, “Spikingjelly: An open- source machine learning infrastructure platform for spike-based intelligence,”Science Advances, vol. 9, p. eadi1480, 2023

  46. [49]

    Neurophysiological investigation of the basis of the fmri signal,

    N. K. Logothetis, J. Pauls, M. Augath, T. Trinath, and A. Oel- termann, “Neurophysiological investigation of the basis of the fmri signal,”Nature, vol. 412, no. 6843, pp. 150–157, 2001

  47. [50]

    Neural population coding: combining insights from microscopic and mass signals,

    S. Panzeri, J. H. Macke, J. Gross, and C. Kayser, “Neural population coding: combining insights from microscopic and mass signals,”Trends in Cognitive Sciences, vol. 19, no. 3, pp. 162–172, 2015

  48. [51]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gellyet al., “An image is worth 16x16 words: Transformers for image recognition at scale,” inInternational Conference on Learning Representations

  49. [52]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inInternational Conference on Machine Learning, 2021, pp. 8748–8763