MindAligner (ICML 2025) demonstrated that an explicit, voxel-level functional alignment — a low-rank linear Brain Transfer Matrix (BTM) learned from roughly one hour of 7-tesla (7T) fMRI per subject, with a cross-stimulus neural mapper conditioned on stimulus differences — can project a new subject's brain signals onto a known subject and enable reuse of pretrained visual decoding models. This was shown only on the Natural Scenes Dataset (NSD): 7T scans, natural photographs, human subjects, and pairwise new-subject-to-one-target mappings. The open problem is whether this explicit alignment mechanism generalizes beyond that single paradigm: (i) to substantially different stimulus domains, experimental paradigms, imaging protocols, or modalities; (ii) to substantially different brains, across species such as human versus macaque; and (iii) from pairwise mappings to a coherent multi-subject reference alignment (atlas) that aggregates many subjects' explicit mappings while preserving interpretability and transferability. Relatedly, the minimal-data and scan-quality limits of the approach remain to be characterized.

Background

Brain (visual) decoding seeks to reconstruct or identify what a person perceived — typically a viewed image — from functional magnetic resonance imaging (fMRI) signals, which measure brain activity indirectly via blood-oxygenation changes with millimeter-scale spatial resolution. A fundamental obstacle is inter-individual variability: the folding anatomy and the fine-grained functional layout of visual cortex differ substantially across people, so a decoder trained on one subject performs poorly on another, and per-subject training is expensive because fMRI data are scarce (hours of scanning are costly). The Natural Scenes Dataset (NSD, Allen et al., Nature Neuroscience 2022) partially mitigates data scarcity: it contains 7-tesla fMRI from 8 subjects who each viewed thousands of overlapping natural photographs drawn from MSCOCO, a large image-captioning dataset, and has become the standard benchmark for data-limited decoding.

Two families of approaches address cross-subject variability. Implicit alignment methods learn subject-specific parameters that map each subject's signals into a shared latent space assumed to capture common cognition; they avoid explicit voxel-to-voxel correspondence but, as MindAligner's authors argue, often achieve insufficient alignment when fMRI data are limited. Explicit functional alignment instead learns a direct mapping between subjects' voxels — the classical example being hyperalignment, which rotates one subject's voxel space into another's so that corresponding functional areas match. MindAligner (Dai et al., ICML 2025) revived explicit alignment for modern image-reconstruction decoders: it learns a Brain Transfer Matrix (BTM) — decomposed into two low-rank linear layers so it can be fit from limited data — that projects a new subject's voxel-level signals onto one of the known training subjects, after which any pretrained decoder for that target subject can be reused unchanged. Because strictly paired data (two subjects viewing identical images) are scarce, a Brain Functional Alignment module performs soft alignment between responses to different but similar images, using a cross-stimulus neural mapper that transforms fMRI signals conditioned on the difference between the two stimuli, trained with multi-level losses (signal-level reconstruction plus latent alignment guided by visual-semantic similarity). With a single NSD session (about one hour) per subject, this outperformed implicit-alignment and per-subject baselines on image reconstruction and retrieval.

The demonstration is, however, confined to one regime: 7T NSD recordings, human adults, natural photographic stimuli, one experimental paradigm, and pairwise mappings of each new subject to a single chosen target subject. The authors' impact statement explicitly points beyond this regime — the approach "holds significant potential for enabling alignment across diverse data formats and uncovering commonalities in brain organization across species, such as between humans and monkeys" and "could play a pivotal role in advancing the creation of unified brain atlases" — while the capacity of a linear low-rank mapping to capture such correspondences, the behavior under very different stimuli or protocols, and the aggregation of many per-subject explicit mappings into one consistent multi-subject reference all remain undemonstrated. The open problem is the generality of the explicit-alignment mechanism itself, as an empirical question about method capability, not merely the engineering of one further decoder.

Problem Statement

Establish, by explicit construction and empirical demonstration, whether voxel-level explicit functional alignment of the MindAligner class — a low-rank linear Brain Transfer Matrix fitted from limited data, optionally with a cross-stimulus mapper conditioned on stimulus differences — is a general mechanism for cross-subject brain decoding, by extending it beyond the 7T NSD natural-image paradigm in which it was demonstrated. A complete answer addresses three linked facets of the same generality question:

  1. Domain and protocol generalization. Demonstrate and characterize the approach on at least one stimulus domain and experimental paradigm substantially different from NSD natural photographs (e.g., non-naturalistic stimuli, a different experimental design) and under imaging conditions differing in field strength, resolution, or acquisition protocol, reporting alignment quality and downstream decoding against appropriate per-subject and implicit-alignment baselines.

  2. Cross-species correspondence. Test whether explicit voxel-level functional alignment can capture correspondences between substantially different brains — e.g., human versus macaque — including whether the linear low-rank form is expressive enough, and if not, providing a controlled characterization of where and why it fails relative to adequately expressive alternatives.

  3. Aggregated multi-subject reference. Go beyond pairwise subject-to-one-target mappings: aggregate explicit mappings from many subjects into a single coherent multi-subject reference alignment (atlas) that preserves voxel-level interpretability, remains consistent across subjects (mappings compose coherently rather than contradicting one another), and transfers to new subjects, with validation on downstream decoding and functional-analysis tasks.

  4. Data and quality limits (related characterization; not one of the three facets). Beyond the facets above, characterize the minimal amount of data and the scan-quality range under which the explicit alignment remains reliable, beyond the single demonstrated point of one hour of high-quality 7T data.

The facets share one scientific objective — the generality of the explicit-alignment mechanism — and an answer must make explicit how the demonstrated (or refuted) capability in each facet bears on that common objective. Progress on any single facet with the others left open is a legitimate partial result; a claim of having solved the problem as a whole requires all three facets — the data-and-quality characterization of item 4 is complementary context rather than an additional requirement.

References

  1. Yuqin Dai, Zhouheng Yao, Chunfeng Song, Qihao Zheng, Weijian Mai, Kunyu Peng, Shuai Lu, Wanli Ouyang, Jian Yang, Jiamin Wu. MindAligner: Explicit Brain Functional Alignment for Cross-Subject Visual Decoding from Limited fMRI Data. ICML 2025. arXiv:2502.05034, DOI 10.48550/arXiv.2502.05034, https://arxiv.org/abs/2502.05034
  2. Emily J. Allen, Ghislain St-Yves, Yihan Wu, Jesse L. Breedlove, Jacob S. Prince, Logan T. Dowdle, Matthias Nau, Brad Caron, Franco Pestilli, Ian Charest, J. Benjamin Hutchinson, Thomas Naselaris, Kendrick Kay. A massive 7T fMRI dataset to bridge cognitive neuroscience and artificial intelligence. Nature Neuroscience 25(1):116–126, 2022. DOI 10.1038/s41593-021-00962-x, https://doi.org/10.1038/s41593-021-00962-x
  3. Shumeng Li, Jintao Guo, Jian Zhang, Yulin Zhou, Luyang Cao, Yinghuan Shi. Duala: Dual-Level Alignment of Subjects and Stimuli for Cross-Subject fMRI Decoding. arXiv:2603.07625, 2026, https://arxiv.org/abs/2603.07625
  4. Jintao Guo, Lin Wang, Shumeng Li, Jian Zhang, Yulin Zhou, Luyang Cao, Hairong Zheng, Yinghuan Shi. StableMind: Source-Free Cross-Subject fMRI Decoding with Regularized Adaptation. arXiv:2605.02586, 2026, https://arxiv.org/abs/2605.02586
  5. Jiaxiang Liu, Jiawei Du, Xupeng Chen, Guoqi Li, Jiang Cai, Simon Fong, Mingkun Xu. MindAdapter: Few-Shot Parameter-Efficient Residual Calibration of Cross-Subject Brain-to-Visual Decoding Models. KDD 2026 (AI4Sciences Track). arXiv:2605.24679, https://arxiv.org/abs/2605.24679
  6. Mu Nan, Muquan Yu, Weijian Mai, Jacob S. Prince, Hossein Adeli, Rui Zhang, Jiahang Cao, Benjamin Becker, John A. Pyles, Margaret M. Henderson, Chunfeng Song, Nikolaus Kriegeskorte, Michael J. Tarr, Xiaoqing Hu, Andrew F. Luo. Meta-learning In-Context Enables Training-Free Cross Subject Brain Decoding. CVPR 2026. arXiv:2604.08537, 2026, https://arxiv.org/abs/2604.08537
  7. Weijian Mai, Mu Nan, Yu Zhu, Jiahang Cao, Rui Zhang, Yuqin Dai, Chunfeng Song, Andrew F. Luo, Jiamin Wu. NeuroFlow: Toward Unified Visual Encoding and Decoding from Neural Activity. CVPR 2026. arXiv:2604.09817, 2026, https://arxiv.org/abs/2604.09817
  8. Nils Leutenegger. Cross-Species RSA Reveals Conserved Early Visual Alignment but Divergent Higher-Area Rankings Across Human fMRI and Macaque Electrophysiology. arXiv:2605.22401, 2026 (v2 revised 2026-08-14 with a correction to the predictive-coding and STDP conditions), https://arxiv.org/abs/2605.22401