A 2025 reanalysis by Shirakawa, Nagano, Tanaka, Cheng, and Kamitani raised five documented methodological concerns about the mental-image-reconstruction study of Koide-Majima, Nishimoto, and Majima (Neural Networks, 2024) — selective reporting, circular evaluation metrics, insufficiently controlled baselines, an effectively inert Bayesian sampling component, and an unsubstantiated 'Bayesian' novelty claim — and left a set of unresolved questions: whether a methodologically valid, reproducible re-evaluation based on the publicly released decoded fMRI features, released reconstruction code, and pretrained generator, with independent perceptual validation and properly controlled ablations, would uphold or overturn the study's claimed advantages, and whether a genuine Bayesian treatment that leverages the latent posterior — posterior-based estimators or uncertainty quantification — yields any practical reconstruction advantage over single-point MAP-style optimization; the reanalysis's appendix already establishes that latent-space and image-space maximum-a-posteriori estimates provably need not correspond under a nonlinear generator. As of August 2026 no later work has adjudicated these questions: the original authors have not released a complete reproducibility package or published a reply, and follow-up literature generalizes the credibility problem without closing this specific core.

Background

Reconstructing seen or imagined images from human brain activity (measured with functional magnetic resonance imaging, fMRI) is a prominent brain-decoding task. The influential framework of Shen, Horikawa, Majima, and Kamitani (2019) works in two steps: a feature-decoding model maps fMRI activity to features of a deep neural network (DNN) trained on image classification, and an image is then synthesized by optimizing the latent vector of a pretrained deep generative network so that the DNN features of the generated image match the features predicted from the brain; this optimization-based pipeline is often called iCNN-style reconstruction (iCNN abbreviates the inversion of a CNN, since the synthesis optimizes a network input so that the CNN's features match target features). Koide-Majima, Nishimoto, and Majima (Neural Networks 170:349–363, 2024; doi:10.1016/j.neunet.2023.11.024) extended this framework with two components and reported superior reconstruction of imagined images: (i) stochastic gradient Langevin dynamics (SGLD), a Markov chain Monte Carlo method that adds Gaussian noise to gradient updates so that, in principle, the iterates sample from a posterior distribution over latent codes rather than converge to a single optimum; and (ii) image features from contrastive language–image pretraining (CLIP), a network trained to align images with natural-language descriptions, intended to add semantic information to the reconstruction targets. The authors framed the method as 'Bayesian estimation' and the study was widely promoted through institutional press releases. A reanalysis by Shirakawa, Nagano, Tanaka, Cheng, and Kamitani (arXiv:2511.07960, 2025; code at github.com/KamitaniLab/repro_mental_image_recon) re-ran the released code and concluded that: (1) the published results reflect selective reporting of the best examples at several levels; (2) the evaluation metrics were circular, i.e. computed in feature spaces related to those optimized during reconstruction, a known inflation risk in systems neuroscience (Kriegeskorte et al. 2009); (3) under fair baseline comparisons, the SGLD and CLIP components showed no discernible advantage; (4) the SGLD implementation is functionally inert — its step size and temperature make the iterates collapse to the deterministic optimization result, so only a single point estimate is effectively used; and (5) the 'Bayesian' framing is nominal: maximum a posteriori (MAP) estimation, the point estimate that maximizes the posterior density, does not by itself confer Bayesian advantages such as uncertainty quantification, and the prior iCNN framework can equally be reinterpreted as MAP estimation. A mathematical appendix adds a geometric subtlety: for a nonlinear generator gg mapping latent codes z\mathbf{z} to images, the change-of-variables (Jacobian) factor distorts densities, so the image-space MAP estimate and the pushed-forward latent-space MAP estimate need not coincide, sMAP≠g(zMAP)\mathbf{s}_{\mathrm{MAP}}\neq g(\mathbf{z}_{\mathrm{MAP}}), with a log-normal distribution as the standard one-dimensional example; conversely, plain latent optimization is equivalent to MAP estimation on the generator manifold (the set of images reachable as g(z)g(\mathbf{z})) under a specific Jacobian-based prior. The reanalysis ends with recommendations for research practice in the field rather than with resolved questions. Related work by the same group — 'Spurious reconstruction from brain activity' (Shirakawa et al., Neural Networks 190:107515, 2025) and a 2026 preprint on 'phantom evidence' in generative-AI-based science — generalizes the credibility problem but does not answer the method-specific questions below.

Problem Statement

Determine, by methodologically valid and reproducible re-evaluation using the publicly released decoded fMRI features, the released reconstruction code, and the pretrained generator, whether the two claimed innovations of Koide-Majima, Nishimoto, and Majima (Neural Networks 170:349–363, 2024) — (i) latent-space posterior sampling via stochastic gradient Langevin dynamics (SGLD) marketed as 'Bayesian estimation' and (ii) CLIP-based semantic features — genuinely improve the reconstruction of mental images from human brain activity relative to the deterministic iCNN baseline (the reanalysis's 'No SGLD + No CLIP' ablation condition, essentially equivalent to Shen et al., 2019), when assessed with controlled ablation experiments that hold all other implementation details fixed (feature decoding, generator, loss terms, update schedules) so the unique contributions of the sampling scheme and of the semantic features are isolated, across the released subjects and imagery stimuli, and with evaluation independent of the feature spaces optimized during reconstruction, via human perceptual validation or evaluation metrics independent of the optimization objective; and establish whether a genuine Bayesian treatment that leverages the latent posterior — e.g., sampling with effective stochasticity, posterior-mean estimation sPM=E[g(z)]≈1T∑i=1Tg(zi)\mathbf{s}_{\mathrm{PM}}=\mathbb{E}[g(\mathbf{z})]\approx\frac{1}{T}\sum_{i=1}^{T}g(\mathbf{z}_{i}), or uncertainty quantification — yields any practical reconstruction advantage over single-point MAP-style optimization for this task, or showing that it does not. The reanalysis's appendix result that sMAP≠g(zMAP)\mathbf{s}_{\mathrm{MAP}} \neq g(\mathbf{z}_{\mathrm{MAP}}) for nonlinear generators gg — with equality when gg is linear or the Jacobian determinant is constant — is established background for this question rather than a deliverable.

References

  1. Ken Shirakawa, Yoshihiro Nagano, Misato Tanaka, Fan L. Cheng, Yukiyasu Kamitani (2025). Advancing credibility and transparency in brain-to-image reconstruction research: Reanalysis of Koide-Majima, Nishimoto, and Majima (Neural Networks, 2024). arXiv:2511.07960, DOI 10.48550/arXiv.2511.07960. https://arxiv.org/abs/2511.07960
  2. Naoko Koide-Majima, Shinji Nishimoto, Kei Majima (2024). Mental image reconstruction from human brain activity: Neural decoding of mental imagery via deep neural network-based Bayesian estimation. Neural Networks 170, 349–363. DOI 10.1016/j.neunet.2023.11.024. https://doi.org/10.1016/j.neunet.2023.11.024
  3. Guohua Shen, Tomoyasu Horikawa, Kei Majima, Yukiyasu Kamitani (2019). Deep image reconstruction from human brain activity. PLoS Computational Biology 15(1), e1006633. DOI 10.1371/journal.pcbi.1006633. https://doi.org/10.1371/journal.pcbi.1006633
  4. Ken Shirakawa, Yoshihiro Nagano, Misato Tanaka, Shuntaro C. Aoki, Yusuke Muraki, Kei Majima, Yukiyasu Kamitani (2025). Spurious reconstruction from brain activity. Neural Networks 190, 107515. DOI 10.1016/j.neunet.2025.107515. https://doi.org/10.1016/j.neunet.2025.107515
  5. Yukiyasu Kamitani, Misato Tanaka, Ken Shirakawa (2025). Visual image reconstruction from brain activity via latent representation. Annual Review of Vision Science 11, 611–634. DOI 10.1146/annurev-vision-110423-023616. https://doi.org/10.1146/annurev-vision-110423-023616
  6. Yukiyasu Kamitani, Ken Shirakawa (2026). Phantom Evidence: How and Why Generative AI Manufactures False Positives in Science. arXiv:2607.25991. https://arxiv.org/abs/2607.25991
  7. Nikolaus Kriegeskorte, W Kyle Simmons, Patrick S F Bellgowan, Chris I Baker (2009). Circular analysis in systems neuroscience: the dangers of double dipping. Nature Neuroscience 12, 535–540. DOI 10.1038/nn.2303. https://doi.org/10.1038/nn.2303
  8. Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever (2021). Learning Transferable Visual Models From Natural Language Supervision. arXiv:2103.00020. https://arxiv.org/abs/2103.00020