Platonic Representations in the Human Brain: Unsupervised Recovery of Universal Geometry
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: I'm Ines.
Marcus: And I'm Marcus, here with Yuki.
Ines: We're looking at the paper: "Platonic Representations in the Human Brain".
Marcus: Here's the gist of the paper. The Strong Platonic Representation Hypothesis suggests that representational convergence in artificial neural networks can be harnessed constructively: embeddings can be translated across models through a universal latent space without paired data.
Ines: This paper investigates whether an analogous geometry can be recovered across human brains using fMRI data.
Marcus: Problem Setup The core problem addressed is testing whether subject-specific fMRI embedding spaces can be translated across subjects using only their intrinsic representational geometry.
Ines: Subjects observe disjoint image sets independently sampled from a distribution D, with no shared images or paired cross-subject correspondences available for learning the translations.
Title and authors: Ines: So this paper is called Platonic Representations in the Human Brain: Unsupervised Recovery of Universal Geometry. It suggests that the way our brains encode things has a universal shape that we can actually recover using just fMRI data from natural scenes, without needing any paired data between subjects.
Marcus: What I’m hearing is they are testing if these subject-specific brain embeddings can be translated across different people using only the intrinsic geometry of those responses, without ever seeing images that belong to both subjects at the same time.
Yuki: It sounds like they are asking if there's a common mathematical structure underlying how different people see and encode natural scenes, and this paper proposes a way to find that structure directly from brain scans.
Ines: Exactly. The core idea is that you can learn embeddings for each subject separately, just by showing them many repeated images, and then use those learned shapes to figure out the relationship between subjects without needing any shared images or training data linking them together initially.
Marcus: They propose a self-supervised encoder that learns these subject-specific fMRI embeddings from repeated stimulus presentations, using voxel reliability weighting and a low-dimensional linear projection with principal component analysis.
Yuki: And then they move on to the next step, which is estimating pairwise brain-to-brain translation estimates called orthogonal rotations, based on geometry derived pseudocorrespondences.
Ines: That’s where things get interesting because it sounds like they are using these geometry-derived pseudocorrespondences to estimate those orthogonal rotations between subjects without having any actual paired cross-subject data for the learning step.
Marcus: They refine those pairwise rotation estimates iteratively, similar to a method called Iterative Closest Point, which is a way to align shapes when you don't have perfect matches.
Yuki: Then they take those pairwise rotations and synchronize them into a single shared latent space by solving an orthogonal synchronization problem over the dimension d, which recovers one rotation per subject mapping all embeddings into a common coordinate system.
Ines: So what they’re doing is taking these independently learned subject spaces and making sure they fit together in one big, shared coordinate system by finding those pairwise rotations for every single person involved.
Marcus: The evaluation was done on the Natural Scenes Dataset, using subject-specific non-shared images to learn the encoders and brain-to-brain translations. They measured performance across two settings: within subjects and across subjects.
Yuki: For within subjects, they found that the embeddings are stable, with an average rank of five point twenty-eight which suggests that even if we look at the same person’s data over time, their representation stays consistent.
Ines: But when they looked across subjects, the synchronization improved retrieval metrics from a score of two point forty-seven to one point ninety-seven and R at one from zero point seventy nine to zero point eighty three. That improvement tells us that aligning these spaces helps them find the right images for each other much better.
Title and authors: Marcus: The results support the existence of an approximately isometric shared neural geometry recoverable directly from fMRI data, which means subject-specific spaces are compatible with a common coordinate system, according to page two of this paper <ref:2605.20496#pg1>.
Yuki: That’s significant because it suggests that we don't need shared stimuli to see if two brains are fundamentally organized in a similar way.
Ines: It has practical implications for cross-subject neural modeling because it means data and encoders learned in one subject can be transferred to another, reducing the need for extensive subject-specific calibration when you move from fMRI to image decoding or vice versa.
Marcus: But they also made a distinction about model brain alignment, suggesting that model features aren't necessarily related by the same near-isometric transformations across subjects.
Yuki: That’s important because it separates the idea of geometric equivalence from what actually happens when you train an AI model on top of those brain representations.
Ines: So, to wrap up this Platonic Representations in the Human Brain: Unsupervised Recovery of Universal Geometry paper, they conclude that subject-specific embeddings learned from brain data alone can be translated across individuals using unsupervised orthogonal rotations without needing paired cross-subject samples.
Marcus: They also found that synchronizing those pairwise rotations into a single shared latent space further indicates that the recovered subject spaces are approximately isometric.
Yuki: That means we can use this method to build a common framework for understanding encoding and pattern component analysis across different subjects, which is what page two of this paper suggests <ref:2605.20496#pg1>.
Ines: It’s a big step because it moves us toward being able to transfer knowledge learned in one person’s brain to another person’s without needing their specific data again.
Marcus: For future work, the authors flag that the method relies on high-quality fMRI data and that the unpaired translation procedure can be sensitive depending on how you initialize those rotations.
Yuki: So while they found this geometry in this setting, they also noted that similar isometries might not hold if we test it in lower signal-to-noise modalities or with stimulus distributions further away from what was used in the Natural Scenes Dataset.
Ines: It’s a necessary caution, because if you apply these findings to other types of data or different visual tasks, that assumed near-isometric relationship might break down.
Marcus: In short, they’ve shown that subject-specific embeddings learned from brain data alone can be translated across individuals using unsupervised orthogonal rotations without paired cross-subject samples.
Yuki: It really suggests a level of universality in the organization of visual processing that is independent of the specific training data used for each individual.
Marcus: For future work, they flag that the method relies on high-quality fMRI data and that the unpaired translation procedure can be sensitive depending on how you initialize those rotations.
Ines: It’s a necessary caution, because if you apply these findings to other types of data or different visual tasks, that assumed near-isometric relationship might break down.
The paper's summary: Ines: So, we're looking at how this paper summarizes its main finding for us. It basically says that even though each person’s brain has its own unique map of what they see, there’s an underlying geometric structure common to all of them.
Marcus: Right. They’re not talking about the actual images themselves, but the mathematical shape of how those images get encoded in the brain scans. They learned subject-specific embeddings from repeated views and then figured out how to translate those across subjects without any paired data between them.
Ines: That’s the core claim: you can map these independent subject spaces into one shared latent space just by figuring out the right orthogonal rotations between them using geometry derived pseudocorrespondences. It means that the distance relationships in one person's brain representation are preserved when you translate it to another person's brain representation.
Marcus: Exactly. The numbers they found were pretty compelling for this part of the math. They showed that within a single subject, the embeddings stay stable, with an average rank of five point two eight, which means their representation doesn't drift much over time.
Ines: But across subjects, when they synchronized those spaces by solving that orthogonal synchronization problem over the dimension d, retrieval metrics got better. They went from two point four seven to one point nine seven for cross-subject retrieval R at one going from zero point seventy nine to zero point eighty three.
Marcus: And that improvement suggests that aligning these spaces actually helps you find the right images much better when you're trying to match brain activity across different people. It supports the idea that subject-specific spaces are compatible with a common coordinate system.
Ines: This has big implications for how we think about neural modeling, Marcus. If you can translate data and encoders learned in one person to another, you don't need to do all that extensive subject-specific calibration every single time you move from fMRI data to image decoding or vice versa.
Marcus: That’s the practical part. It suggests that the underlying structure of visual perception is more universal than we thought, even if the specific pathways differ between individuals. But they did put a stop to it too.
Ines: They made a distinction about model brain alignment, which is important because it suggests that just because neural embeddings are geometrically related doesn't mean the features in an actual AI model are related by those exact same near-isometric transformations across subjects.
Marcus: So, the geometry is there, but it's not a one-to-one mapping for what we see in deep learning models. They’re looking at different things here—the brain structure versus the learned weights of an artificial system.
Ines: And they did flag some limitations too. The method relies on high quality fMRI data, and the way they do that unpaired translation procedure can be sensitive depending on how you initialize those rotations.
Marcus: True. It’s not a plug-and-play solution for any dataset you throw at it without careful tuning of the starting point for those transformations. But overall, they've shown that subject-specific embeddings learned from brain data alone can be translated across individuals using unsupervised orthogonal rotations without needing paired cross-subject samples.
Ines: It really does suggest a level of universality in the organization of visual processing that is independent of the specific training data used for each individual.
The paper's improvements: Tom: So, we’re talking about how the authors suggest they can actually make this method better than what they first proposed. It sounds like they’re looking for ways to refine those rotations and synchronization steps.
Ines: Right. They mention that the method works best when you use high quality fMRI data, which is a big caveat for us, but they also point out that the unpaired translation procedure can be sensitive depending on how you set up those initial rotations.
Marcus: So what they’re saying is that if you want to use this across different datasets or modalities, like maybe lower signal-to-noise data, you need to be extra careful with how you start the process. You can't just throw any initialization at it and expect a good result.
Ines: That makes sense because they’re working with these geometry derived pseudocorrespondences, and if those initial geometric guesses are off by even a little bit, the whole synchronization into that shared latent space might get messy. They need robust methods for getting that starting point right.
Marcus: And they also hint at the future work there. They suggest testing whether this kind of isometry—this near-perfect shared geometry—actually holds up if you move to different stimulus distributions or even lower signal-to-noise modalities than what they used on the Natural Scenes Dataset.
Ines: That’s a big question for the biology side. If this universal geometry we found only works well for natural scenes, does it mean that when we look at human brains responding to something completely different, like a medical scan or a synthetic visual task, this specific kind of shared structure might disappear?
Marcus: And that brings us back to the model brain alignment part. They’ve shown this geometric equivalence in the neural embeddings, but they’re still separating that from what happens when you look at actual AI models. The geometry is one thing, but how an AI learns its weights isn't necessarily governed by these same near-isometric transformations across subjects.
Ines: Exactly. So the implication is that while we can transfer data and encoders between people using this geometric alignment, we shouldn't assume that the underlying functional mapping in a machine learning model will follow those exact same rules when it’s being compared across different subjects or tasks.
Marcus: It means the utility of this method is really focused on cross-subject neural modeling, but it doesn't claim to solve the whole problem of how AI learns representations from scratch. They’re showing a tool for aligning brain data, not a complete theory of general learning transfer.
Ines: So, the takeaway is that we have a way to see if two brains are fundamentally organized in a similar way just by looking at their response geometry, but we have to be cautious about applying that finding outside of natural scenes and remember the difference between neural geometry and artificial model weights.
Conclusion: Ines: So we’re wrapping up this look at "Platonic Representations in the Human Brain: Unsupervised Recovery of Universal Geometry." It really boils down to showing that subject-specific embeddings learned from fMRI data can be translated between individuals using just geometry derived pseudocorrespondences, without needing any paired cross-subject images.
Marcus: Right. The big deal is that this suggests an approximately isometric shared neural geometry is recoverable directly from the fMRI data, meaning the spaces learned for one person are compatible with a common coordinate system across everyone.
Yuki: From a population perspective, it’s interesting because it suggests that even though individual brain responses vary, there's a deep structural similarity in how those responses are organized across different people. It supports the idea that there's some underlying universal organization to visual processing.
Ines: That’s what I was thinking. It means we can actually use data and encoders learned in one subject for another person, which is a massive practical thing if we ever want to do cross-subject neural modeling without needing tons of extra calibration time.
Marcus: It definitely opens up the door for more efficient transfer learning in neuroscience, but remember they also pointed out that model brain alignment isn't necessarily related by those same near-isometric transformations across subjects. That’s a key distinction we have to keep in mind.
Yuki: I just wonder if this universal geometry holds up when we test it on stimulus distributions that are really different from the natural scenes they used for the Natural Scenes Dataset, because that could be a limiting factor for its broader application in population genetics.
Ines: True. That’s the caveat—the method relies on high quality fMRI data and is sensitive to initialization of those rotations, so we can't just assume it works everywhere without careful setup. But the core finding about translating subject spaces remains strong even with those limitations.
Marcus: So, to sum up, they’ve shown that subject-specific embeddings learned from brain data alone can be translated across individuals using unsupervised orthogonal rotations without paired cross-subject samples. It's a way to see if two brains are fundamentally organized in a similar way through their response geometry.
Yuki: I think the real impact here is showing that we don't always need shared stimuli to find universal structures in the brain, which is a big conceptual step for population studies.
Ines: Exactly. This work on "Platonic Representations in the Human Brain: Unsupervised Recovery of Universal Geometry" gives us a new tool to test these geometric hypotheses directly from brain data alone.
Marcus: It’s a powerful technique for aligning subject spaces, but we need to keep an eye on those caveats about initialization and the quality of the fMRI signal when we try to apply it in more complex or diverse settings.
University of Barcelona · Cornell University · Bellvitge Institute for Biomedical Research, Spain
q-bio.NC, cs.CV
Submitted: 2026-05-19
Updated: 2026-11-06
Comments: Code available at https://github.com/memory-formation/platonic-representations-fmri
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 80/100
The gist: The Strong Platonic Representation Hypothesis suggests that representational convergence in artificial neural networks can be harnessed constructively…
Key concepts
- Subject-Specific Embeddings
- These are low-dimensional numerical representations of brain activity learned from a single person's fMRI data based on their unique patterns. The researchers used these to see if they could be understood or mapped onto other people's data without any direct comparison between the subjects.
- Orthogonal Rotations (Rs→t)
- These are mathematical transformations used to find the specific way one subject's embedding space relates to another subject's. The method uses these rotations, derived from geometry-based pseudocorrespondences, to align the individual spaces into a single shared coordinate system.
- Isometric Shared Geometry
- This means that the underlying geometric structure of brain representations across different subjects is nearly identical or perfectly preserved. Finding this shared geometry suggests that even though individual brains have unique patterns, they share a fundamental, universal spatial organization.
Terminology
Summary
The Strong Platonic Representation Hypothesis suggests that representational convergence in artificial neural networks can be harnessed constructively: embeddings can be translated across models through a universal latent space without paired data. p. 0 Platonic Representations in the Human Brain:
Unsupervised Recovery of Universal Geometry
Pablo Marcos-Manchón1,2 Rishi Jha3 Lluís Fuentemilla1,2,4
1Department of Cognition, Development and Education Psychology, University of Barcelona
2
Institute of Neurosciences, University of Barcelona
3Department of Computer Science, Cornell University
4Bellvitge Institute for Biomedical Research, Spain
Abstract
The Strong Platonic Representation Hypothesis suggests that representational convergence in artificial neural networks can be harnessed constructively: embeddings
can be translated across models through a universal latent space without paired data. We ask whether an analogous geometry can be recovered across human brains. We ask whether an analogous geometry can be recovered across human brains. Using fMRI data from the Natural Scenes Dataset, we propose a self-supervised
encoder that learns subject-specific embeddings from brain data alone by exploiting
repeated stimulus presentations.Where the paper says it
Where the paper says it
Problem Setup
The core problem addressed is testing whether subject-specific fMRI embedding spaces can be translated across subjects using only their intrinsic representational geometry. Subjects observe disjoint image sets independently sampled from a distribution D, with no shared images or paired cross-subject correspondences available for learning the translations. p. 1 Subjects observe disjoint image sets independently sampled from D. No shared images or paired
cross-subject correspondences are available for learning the translations. X(s) ∈ R
ns×vs denotes the subject’s neural activity matrix, where rows correspond
to image presentations and columns to subject-specific voxel responses. Our goal is to test whether
these independently observed fMRI responses can be mapped into a shared latent space Z using only
their intrinsic representational geometry, without paired cross-subject data for learning the mappings.Where the paper says it
Where the paper says it
Methodology
The method proceeds in three stages:
-
A self-supervised encoder learns subject-specific fMRI embeddings from repeated stimulus presentations. p. 0
This involves voxel reliability weighting to mitigate temporal drifts and a low-dimensional linear projection using PCA. p. 2Where the paper says it
We ask whether an analogous geometry can be recovered across human brains. Using fMRI data from the Natural Scenes Dataset, we propose a self-supervised encoder that learns subject-specific embeddings from brain data alone by exploiting repeated stimulus presentations.
Where the paper says it
(A) Subject encoder. For each subject, fMRI responses are mapped into a low-dimensional embedding space using voxel reliability weighting, PCA, and multi-view CCA (MCCA), followed by a residual nonlinear refinement trained from repeated stimulus presentations.
-
Pairwise brain-to-brain translation estimates unsupervised orthogonal rotations Rs→t from geometry-derived pseudocorrespondences. p. 2
This is refined iteratively using a method similar to Iterative Closest Point [26]. p. 3Where the paper says it
(B) Pairwise brain-to-brain translation. Independently learned subject embeddings are translated between subject pairs by estimating orthogonal rotations Rs→t from geometry-derived pseudocorrespondences.
Where the paper says it
These pairs define the initial orthogonal Procrustes problem: R (0) s→t = arg min R∈O(d) ∥Z (s)R − Z˜(t) ∥ 2 F . (8) We refine the translation iteratively using an approach similar to Iterative Closest Point [26].
-
Pairwise rotations are synchronized into a single shared latent space by solving an orthogonal synchronization problem over O(d). p. 4
This recovers one transformation Rs per subject mapping all embeddings into a common coordinate system. p. 6Where the paper says it
3 Shared latent space construction Given pairwise translations {Rs→t} S s,t=1, we construct a shared latent space Z by solving an orthogonal synchronization problem over O(d) [27, 28]. The goal is to recover one transformation Rs ∈ O(d) per subject such that Rs→t ≈ RsR⊤ t , placing all subjects in a common coordinate system, i.
Where the paper says it
We therefore integrate the pairwise rotations into a all-subject shared latent space by recovering one orthogonal transformation Rs per subject through global synchronization (Subsection 3. 3), mapping all embeddings into a common coordinate system (Fig.
Experimental Evaluation
The method was evaluated on the Natural Scenes Dataset (NSD) [18], using subject-specific, non-shared images to learn encoders and brain-to-brain translations. Performance is measured across two settings: within subjects and across subjects. p. 4 We evaluate retrieval in two settings. Within subjects, embeddings from one repetition retrieve the
matching image from another repetition among 515 candidates, averaged across repetition pairs. In
general our embeddings are stable: with an average rank of 5. 28, within-subject retrieval works well
across subjects. We show that these independently learned spaces
can be translated across subjects using unsupervised orthogonal rotations, without
paired cross-subject samples or intermediate model representations. Synchronizing
pairwise rotations into a single shared latent space further improves cross-subject
retrieval, indicating that subject-specific spaces are mutually compatible with a
common coordinate system.Where the paper says it
Where the paper says it
Where the paper says it
Discussion
The results support the existence of an approximately isometric shared neural geometry recoverable directly from fMRI data. This suggests that subject-specific spaces are mutually compatible with a common coordinate system. p. 0 We show that these independently learned spaces
can be translated across subjects using unsupervised orthogonal rotations, without
paired cross-subject samples or intermediate model representations. Synchronizing
pairwise rotations into a single shared latent space further improves cross-subject
retrieval, indicating that subject-specific spaces are mutually compatible with a
common coordinate system. This shared neural geometry has practical implications for cross-subject neural modeling. If subjectspecific spaces can be placed into a common coordinate system, then data, encoders, and decoders
learned in one subject may become transferable to another, reducing the need for extensive subjectspecific calibration. Our model–brain results further separate decodability from geometric equivalence. Model features
predict neural embeddings under supervised linear mappings, but are not related to them by the
same near-isometric transformations observed across brains.Where the paper says it
Where the paper says it
Where the paper says it
Related Work
The work situates itself within the context of universal geometry in machine learning, which posits that models converge toward a shared latent structure. It contrasts this with existing shared representations in neuroscience, noting that functional alignment methods typically require shared stimuli. p. 0 However, existing evidence typically depends on shared stimuli,
paired measurements, or external reference spaces. Functional alignment methods for fMRI, such
as hyperalignment, map neural responses into common spaces by using shared stimuli to establish
cross-subject correspondences [14, 15]. Other methods introduce anchors through model-derived feature spaces [16]
or image-to-fMRI encoders [17], while unsupervised alternatives infer correspondences through
structural or distributional matching [45] and optimal transport-based functional alignment [46, 47]. In contrast, we learn subject-specific fMRI embeddings from neural repetitions and test whether they
can be translated by orthogonal rotations learned from unpaired subject-specific images. Our model–brain results further separate decodability from geometric equivalence. Model features
predict neural embeddings under supervised linear mappings, but are not related to them by the
same near-isometric transformations observed across brains.Where the paper says it
Where the paper says it
Where the paper says it
Conclusion and Limitations
The study concludes that subject-specific embeddings learned from brain data alone can be translated across individuals using unsupervised orthogonal rotations without paired cross-subject samples. The synchronization into a shared latent space further indicates that the recovered subject spaces are approximately isometric. p. 8 Synchronizing these rotations into a shared latent space further improved retrieval,
indicating that the pairwise translations are mutually compatible with a common coordinate system. Together, these results suggest that the recovered subject spaces are not only similar, but approximately
isometric. Several limitations remain. The analysis relies on high-quality fMRI data with repeated stimulus
presentations to isolate stable stimulus-related signal from trial-level noise. The analysis relies on high-quality fMRI data with repeated stimulus
presentations to isolate stable stimulus-related signal from trial-level noise. It remains unclear whether
similar isometries will hold in lower-SNR modalities, smaller datasets, or higher-level cognitive
tasks. Future
work should test whether shared neural geometries extend to more distant stimulus distributions, task
contexts, and clinical datasets. Finally, subject-agnostic neural spaces may improve decoding and
transfer learning, but related methods raise neural privacy concerns if applied to sensitive brain data
without consent; future use should require explicit consent, careful data governance, and safeguards
against misuse.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Acknowledgments
This work was supported by the Spanish Ministerio de Ciencia, Innovación y Universidades, which is part of Agencia Estatal de Investigación (AEI), through the project PID2022 - 140426NB (Co-funded by European Regional Development Fund. ERDF, a way to build Europe). p. 9 Acknowledgments and Disclosure of Funding
This work was supported by the Spanish Ministerio de Ciencia, Innovación y Universidades, which is
part of Agencia Estatal de Investigación (AEI), through the project PID2022 - 140426NB (Co-funded
by European Regional Development Fund. ERDF, a way to build Europe). ERDF, a way to build Europe). We thank CERCA
Programme/Generalitat de Catalunya for institutional support. We thank CERCA
Programme/Generalitat de Catalunya for institutional support. was supported by the Digital Life
Initiative Doctoral Fellowship at Cornell Tech.Where the paper says it
Where the paper says it
Where the paper says it
References
[1] Tomas Mikolov, Wen-tau Yih, and Geoffrey Zweig. Linguistic regularities in continuous space word representations. p. 9 References
[1] Tomas Mikolov, Wen-tau Yih, and Geoffrey Zweig. Linguistic regularities in continuous space word
representations.Where the paper says it
[2] Maithra Raghu, Justin Gilmer, Jason Yosinski, and Jascha Sohl-Dickstein. p. 9 Association for
Computational Linguistics, 2013. [2] Maithra Raghu, Justin Gilmer, Jason Yosinski, and Jascha Sohl-Dickstein. [2] Maithra Raghu, Justin Gilmer, Jason Yosinski, and Jascha Sohl-Dickstein. Svcca: singular vector canonical
correlation analysis for deep learning dynamics and interpretability.Where the paper says it
Where the paper says it
[3] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. p. 9 ISBN 9781510860964. [3] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. [3] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network
representations revisited.Where the paper says it
Where the paper says it
[4] Ilia Sucholutsky, Lukas Muttenthaler, Adrian Weller, Andi Peng, Andreea Bobu, Been Kim, Bradley C. p. 9 In International Conference on Machine Learning (ICML), volume 97, pages
3519–3529, 2019. [4] Ilia Sucholutsky, Lukas Muttenthaler, Adrian Weller, Andi Peng, Andreea Bobu, Been Kim, Bradley C. [4] Ilia Sucholutsky, Lukas Muttenthaler, Adrian Weller, Andi Peng, Andreea Bobu, Been Kim, Bradley C. Love, Christopher J Cueva, Erin Grant, Iris Groen, Jascha Achterberg, Joshua B.Where the paper says it
Where the paper says it
[5] Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola. p. 9 id=Hiq7lUh4Yn. [5] Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola. Platonic Representations in the Human Brain:
Unsupervised Recovery of Universal Geometry
Pablo Marcos-Manchón1,2 Rishi Jha3 Lluís Fuentemilla1,2,4
1Department of Cognition, Development and Education Psychology, University of Barcelona
2
Institute of Neurosciences, University of Barcelona
3Department of Computer Science, Cornell University
4Bellvitge Institute for Biomedical Research, Spain
Abstract
The Strong Platonic Representation Hypothesis suggests that representational convergence in artificial neural networks can be harnessed constructively: embeddings
can be translated across models through a universal latent space without paired data. We ask whether an analogous geometry can be recovered across human brains.Where the paper says it
Where the paper says it
[6] Rishi Dev Jha, Collin Zhang, Vitaly Shmatikov, and John Xavier Morris. p. 9 In Proceedings of the 41st International Conference on Machine Learning. [6] Rishi Dev Jha, Collin Zhang, Vitaly Shmatikov, and John Xavier Morris. [6] Rishi Dev Jha, Collin Zhang, Vitaly Shmatikov, and John Xavier Morris. Harnessing the universal geometry
of embeddings.Where the paper says it
Where the paper says it
[7] Guy Dar. p. 9 In Advances in Neural Information Processing Systems, 2025. [7] Guy Dar. [7] Guy Dar. mini-vec2vec: Scaling universal geometry alignment with linear transformations, 2026.Where the paper says it
Where the paper says it
[8] Yamini Bansal, Preetum Nakkiran, and Boaz Barak. p. 9 org/abs/2510. [8] Yamini Bansal, Preetum Nakkiran, and Boaz Barak. [8] Yamini Bansal, Preetum Nakkiran, and Boaz Barak. Revisiting model stitching to compare neural representations.Where the paper says it
Where the paper says it
[9] Jianglin Lu, Hailing Wang, Yi Xu, Yizhou Wang, Kuo Yang, and Yun Fu. p. 9 id=ak06J5jNR4. [9] Jianglin Lu, Hailing Wang, Yi Xu, Yizhou Wang, Kuo Yang, and Yun Fu. [9] Jianglin Lu, Hailing Wang, Yi Xu, Yizhou Wang, Kuo Yang, and Yun Fu. Representation potentials
of foundation models for multimodal alignment: A survey.Where the paper says it
Where the paper says it
[10] Uri Hasson, Yuval Nir, Ifat Levy, Galit Fuhrmann, and Rafael Malach. p. 9 emnlp-main. [10] Uri Hasson, Yuval Nir, Ifat Levy, Galit Fuhrmann, and Rafael Malach. [10] Uri Hasson, Yuval Nir, Ifat Levy, Galit Fuhrmann, and Rafael Malach. Intersubject synchronization of
cortical activity during natural vision.Where the paper says it
Where the paper says it
[11] Samuel A Nastase, Valeria Gazzola, Uri Hasson, and Christian Keysers. p. 9 1126/science. [11] Samuel A Nastase, Valeria Gazzola, Uri Hasson, and Christian Keysers. [11] Samuel A Nastase, Valeria Gazzola, Uri Hasson, and Christian Keysers. Measuring shared responses
across subjects using intersubject correlation.Where the paper says it
Where the paper says it
[12] Nikolaus Kriegeskorte, Marieke Mur, and Peter A. p. 9 1093/scan/nsz037. [12] Nikolaus Kriegeskorte, Marieke Mur, and Peter A. [12] Nikolaus Kriegeskorte, Marieke Mur, and Peter A. Bandettini. Bandettini. Representational similarity analysis –
connecting the branches of systems neuroscience.Where the paper says it
Where the paper says it
Where the paper says it
[13] Baihan Lin and Nikolaus Kriegeskorte. p. 9 3389/neuro. [13] Baihan Lin and Nikolaus Kriegeskorte. [13] Baihan Lin and Nikolaus Kriegeskorte. The topology and geometry of neural representations.Where the paper says it
Where the paper says it
[14] James V. Haxby, J. Swaroop Guntupalli, Andrew C. p. 9 Tenenbaum, Katherine M. Collins, Katherine Hermann, Kerem Oktar, Klaus Greff, Martin N Hebart, Nathan Cloos,
Nikolaus Kriegeskorte, Nori Jacoby, Qiuyi Zhang, Raja Marjieh, Robert Geirhos, Sherol Chen, Simon Kornblith, Sunayana Rane, Talia Konkle, Thomas O’Connell, Thomas Unterthiner, Andrew Kyle
Lampinen, Klaus Robert Muller, Mariya Toneva, and Thomas L.Where the paper says it
[15] J. Swaroop Guntupalli, Michael Hanke, Yaroslav O. Halchenko, Andrew C. p. 9 Tenenbaum, Katherine M. Collins, Katherine Hermann, Kerem Oktar, Klaus Greff, Martin N Hebart, Nathan Cloos,
Nikolaus Kriegeskorte, Nori Jacoby, Qiuyi Zhang, Raja Marjieh, Robert Geirhos, Sherol Chen, Simon Kornblith, Sunayana Rane, Talia Konkle, Thomas O’Connell, Thomas Unterthiner, Andrew Kyle
Lampinen, Klaus Robert Muller, Mariya Toneva, and Thomas L. 1093/scan/nsz037. [12] Nikolaus Kriegeskorte, Marieke Mur, and Peter A.Where the paper says it
Where the paper says it
[16] Qiongyi Zhou, Changde Du, Shengpei Wang, and Huiguang He. p. 9 id=Hiq7lUh4Yn. [5] Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola. 21
DINOv2 ViT-S/14 [33] + Ridge Vision Model Guided 24. 46
CLIP ViT-B/16 [34] + Ridge Vision Model Guided 23.Where the paper says it
Where the paper says it
[17] Navve Wasserman, Roman Beliy, Roy Urbach, and Michal Irani. Functional brain-to-brain transformation without shared stimuli. p. 3 2 Pairwise brain-to-brain translation
To translate embeddings between two subjects, we adapt mini-vec2vec [7] to neural embeddings. Given subjects s and t, we learn an orthogonal transformation Rs→t ∈ O(d) such that Z
(s)Rs→t ≈
Z
(t)
, without paired cross-subject samples during training.Where the paper says it
[18] Emily J. Allen, Ghislain St-Yves, Yihan Wu, Jesse L. Breedlove, Jacob S. Prince, Logan T. Dowdle, Matthias Nau.
[19] Thomas T. p. 9 Tenenbaum, Katherine M. Collins, Katherine Hermann, Kerem Oktar, Klaus Greff, Martin N Hebart, Nathan Cloos,
Nikolaus Kriegeskorte, Nori Jacoby, Qiuyi Zhang, Raja Marjieh, Robert Geirhos, Sherol Chen, Simon Kornblith, Sunayana Rane, Talia Konkle, Thomas O’Connell, Thomas Unterthiner, Andrew Kyle
Lampinen, Klaus Robert Muller, Mariya Toneva, and Thomas L. Several limitations remain. The analysis relies on high-quality fMRI data with repeated stimulus
presentations to isolate stable stimulus-related signal from trial-level noise.Where the paper says it
Where the paper says it
[20] Jacob S Prince, Ian Charest, Jan W Kurzawski, John A Pyles, Michael J Tarr, and Kendrick N Kay. p. 9 In Proceedings of the 41st International Conference on Machine Learning. [6] Rishi Dev Jha, Collin Zhang, Vitaly Shmatikov, and John Xavier Morris. We synchronize pairwise rotations into a single shared latent space, improving cross-subject
retrieval and showing that independently learned subject spaces are mutually compatible
with a common coordinate system. Our results support the existence of an approximately isometric shared neural geometry recoverable
directly from fMRI data, with practical implications for cross-subject neural modeling.Where the paper says it
Where the paper says it
[21] Furkan Ozcelik and Rufin VanRullen. Natural scene reconstruction from fmri signals using generative latent diffusion. p. 0 We ask whether an analogous geometry can be recovered across human brains. Using fMRI data from the Natural Scenes Dataset, we propose a self-supervised
encoder that learns subject-specific embeddings from brain data alone by exploiting
repeated stimulus presentations.Where the paper says it
[22] Paul Steven Scotti, Atmadeep Banerjee, Jimmie Goode, Stepan Shabalin, Alex Nguyen, Cohen Ethan, Aidan James Dempster.
[23] J. R. KETTENRING. Canonical analysis of several sets of variables. p. 8 Several limitations remain. The analysis relies on high-quality fMRI data with repeated stimulus
presentations to isolate stable stimulus-related signal from trial-level noise.Where the paper says it
[24] Arthur Tenenhaus and Michel Tenenhaus. Regularized generalized canonical correlation analysis. p. 2 Let Y ∈ R
rns×dPCA
denote the concatenation of all views. To extract components shared across repetitions, we apply multi-view canonical correlation analysis
(MCCA) to {Yi}
r
i=1 [23, 24].Where the paper says it
[25] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. p. 9 The platonic representation hypothesis. In Proceedings of the 41st International Conference on Machine Learning.Where the paper says it
[26] Paul J. Besl and Neil D. McKay. A method for registration of 3-d shapes. p. 1 This constraint prevents arbitrary
geometric warping and provides a strong test of whether independently learned neural representations
can be translated into a shared coordinate system using geometry alone. 3 Method
We proceed in three stages.Where the paper says it
[27] A Singer. Angular synchronization by eigenvectors and semidefinite programming. p. 4 (11)
We recover a relaxed solution using a spectral method for orthogonal synchronization [27]. We
compute the top-d eigenvectors of B, yielding U ∈ R
Sd×d
, whose blocks {Us}
S
s=1 approximate
the subject-specific transformations.Where the paper says it
[28] Lanhui Wang and Amit Singer. p. 9 id=Hiq7lUh4Yn. [5] Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola. Even for higher-performing subjects,
RSA remains around 0. 6, indicating that robust instance-level identification does not require exact
preservation of the full pairwise geometry.Where the paper says it
Where the paper says it
[29] Tsung-Yi Lin, Michael Maire, Serge J. p. 3 Z¯
lin =
1
r
Xr
i=1
Y Ui
. (3)
We then distill these multi-view projections into a single linear mapping via ridge regression:
W∗ = arg min
W
∥Y W − Z¯
lin∥
2
F + λreg∥W∥
2
F , (4)
yielding Zlin = Y W∗ ∈ R
rns×d
.Where the paper says it
[30] Jacob S. Prince, Ian Charest, Jan W. Kurzawski, John A Pyles. p. 9 In Proceedings of the 41st International Conference on Machine Learning. [6] Rishi Dev Jha, Collin Zhang, Vitaly Shmatikov, and John Xavier Morris.Where the paper says it
[31] Weiran Wang, Raman Arora, Karen Livescu, and Jeff Bilmes. p. 9 id=Hiq7lUh4Yn. [5] Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola. (A) Subject encoder. For each subject, fMRI responses are mapped
into a low-dimensional embedding space using voxel reliability weighting, PCA, and multi-view CCA
(MCCA), followed by a residual nonlinear refinement trained from repeated stimulus presentations.Where the paper says it
Where the paper says it
[32] Galen Andrew, Raman Arora, Jeff Bilmes, and Karen Livescu. p. 9 Tenenbaum, Katherine M. Collins, Katherine Hermann, Kerem Oktar, Klaus Greff, Martin N Hebart, Nathan Cloos,
Nikolaus Kriegeskorte, Nori Jacoby, Qiuyi Zhang, Raja Marjieh, Robert Geirhos, Sherol Chen, Simon Kornblith, Sunayana Rane, Talia Konkle, Thomas O’Connell, Thomas Unterthiner, Andrew Kyle
Lampinen, Klaus Robert Muller, Mariya Toneva, and Thomas L. [2] Maithra Raghu, Justin Gilmer, Jason Yosinski, and Jascha Sohl-Dickstein. Svcca: singular vector canonical
correlation analysis for deep learning dynamics and interpretability.Where the paper says it
Where the paper says it
[33] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec.
[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh. p. 9 In International Conference on Machine Learning (ICML), volume 97, pages
3519–3529, 2019. [4] Ilia Sucholutsky, Lukas Muttenthaler, Adrian Weller, Andi Peng, Andreea Bobu, Been Kim, Bradley C.Where the paper says it
[35] Andreas Peter Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman. p. 9 1093/scan/nsz037. [12] Nikolaus Kriegeskorte, Marieke Mur, and Peter A.Where the paper says it
[36] Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using Siamese BERT-networks. p. 0 Platonic Representations in the Human Brain:
Unsupervised Recovery of Universal Geometry
Pablo Marcos-Manchón1,2 Rishi Jha3 Lluís Fuentemilla1,2,4
1Department of Cognition, Development and Education Psychology, University of Barcelona
2
Institute of Neurosciences, University of Barcelona
3Department of Computer Science, Cornell University
4Bellvitge Institute for Biomedical Research, Spain
Abstract
The Strong Platonic Representation Hypothesis suggests that representational convergence in artificial neural networks can be harnessed constructively: embeddings
can be translated across models through a universal latent space without paired data. We ask whether an analogous geometry can be recovered across human brains.Where the paper says it
[37] Yixuan Li, Jason Yosinski, Jeff Clune, Hod Lipson, and John Hopcroft. p. 9 Association for
Computational Linguistics, 2013. [2] Maithra Raghu, Justin Gilmer, Jason Yosinski, and Jascha Sohl-Dickstein. Learned representations in deep neural networks exhibit
structured geometries that reflect statistical properties of the data. Empirical work shows that models
trained with different architectures, objectives, and modalities often develop similar representational
spaces [1–3, 37].Where the paper says it
Where the paper says it
[38] David Alvarez-Melis and Tommi Jaakkola. Gromov-Wasserstein alignment of word embedding spaces. p. 7 Table 4: Supervised model-to-brain alignment. Last-layer model embeddings are mapped to the
neural embedding space using each subject’s training images and evaluated on the 515 held-out
shared images.Where the paper says it
[39] Edouard Grave, Armand Joulin, and Quentin Berthet. Unsupervised alignment of embeddings with wasserstein procrustes. p. 8 Other methods introduce anchors through model-derived feature spaces [16]
or image-to-fMRI encoders [17], while unsupervised alternatives infer correspondences through
structural or distributional matching [45] and optimal transport-based functional alignment [46, 47]. In contrast, we learn subject-specific fMRI embeddings from neural repetitions and test whether they
can be translated by orthogonal rotations learned from unpaired subject-specific images.Where the paper says it
[40] Luca Moschella, Valentino Maiorca, Marco Fumero, Antonio Norelli, Francesco Locatello.
[41] Guillaume Lample, Alexis Conneau, Marc’Aurelio Ranzato.
[42] Pablo Marcos-Manchón and Lluís Fuentemilla. p. 0 Platonic Representations in the Human Brain:
Unsupervised Recovery of Universal Geometry
Pablo Marcos-Manchón1,2 Rishi Jha3 Lluís Fuentemilla1,2,4
1Department of Cognition, Development and Education Psychology, University of Barcelona
2
Institute of Neurosciences, University of Barcelona
3Department of Computer Science, Cornell University
4Bellvitge Institute for Biomedical Research, Spain
Abstract
The Strong Platonic Representation Hypothesis suggests that representational convergence in artificial neural networks can be harnessed constructively: embeddings
can be translated across models through a universal latent space without paired data. We ask whether an analogous geometry can be recovered across human brains. This could benefit applications such as image-to-fMRI encoding, fMRI-to-image
decoding, and synthetic neural data generation, where paired data are expensive and often unavailable
across subjects. More broadly, shared spaces provide a route for training neural models from
heterogeneous datasets in which different subjects viewed different stimuli.Where the paper says it
Where the paper says it
[43] Janice Chen, Yuan Chang Leong, Christopher J. p. 9 [4] Ilia Sucholutsky, Lukas Muttenthaler, Adrian Weller, Andi Peng, Andreea Bobu, Been Kim, Bradley C. Love, Christopher J Cueva, Erin Grant, Iris Groen, Jascha Achterberg, Joshua B.Where the paper says it
[44] Jörn Diedrichsen and Nikolaus Kriegeskorte. p. 9 Tenenbaum, Katherine M. Collins, Katherine Hermann, Kerem Oktar, Klaus Greff, Martin N Hebart, Nathan Cloos,
Nikolaus Kriegeskorte, Nori Jacoby, Qiuyi Zhang, Raja Marjieh, Robert Geirhos, Sherol Chen, Simon Kornblith, Sunayana Rane, Talia Konkle, Thomas O’Connell, Thomas Unterthiner, Andrew Kyle
Lampinen, Klaus Robert Muller, Mariya Toneva, and Thomas L. These approaches support encoding, decoding, and benchmarking efforts such
as Brain-Score [51–56]. Recent work further suggests that brains and models may share lowdimensional representational axes [57].Where the paper says it
Where the paper says it
[45] Daiki Nakamura, Shizuo Kaji, Ryota Kanai, and Ryusuke Hayashi. Unsupervised method for representation transfer from one brain to another. p. 8 Our method instead recovers a shared
brain space directly from neural data, using model–brain mappings only diagnostically. 6 Discussion
The Strong Platonic Representation Hypothesis [6] posits that the universal latent structure of
representations not only exists but can be harnessed to translate representations from one space to
another without any paired data or model access.Where the paper says it
[46] T. Bazeille, H. Richard, H. Janati, and B. Thirion. Local optimal transport for functional brain template estimation. p. 8 Functional alignment methods such as
hyperalignment map subjects into common spaces, but typically require shared stimuli to establish correspondences [14, 15]. Other methods introduce anchors through model-derived feature spaces [16]
or image-to-fMRI encoders [17], while unsupervised alternatives infer correspondences through
structural or distributional matching [45] and optimal transport-based functional alignment [46, 47].Where the paper says it
[47] Thomas Bazeille, Elizabeth DuPre, Hugo Richard, Jean-Baptiste Poline. p. 9 Tenenbaum, Katherine M. Collins, Katherine Hermann, Kerem Oktar, Klaus Greff, Martin N Hebart, Nathan Cloos,
Nikolaus Kriegeskorte, Nori Jacoby, Qiuyi Zhang, Raja Marjieh, Robert Geirhos, Sherol Chen, Simon Kornblith, Sunayana Rane, Talia Konkle, Thomas O’Connell, Thomas Unterthiner, Andrew Kyle
Lampinen, Klaus Robert Muller, Mariya Toneva, and Thomas L.Where the paper says it
[48] Daniel L. K. Yamins, Ha Hong, Charles F Cadieu, Ethan A Solomon.
[49] Seyed-Mahdi Khaligh-Razavi and Nikolaus Kriegeskorte. p. 9 Tenenbaum, Katherine M. Collins, Katherine Hermann, Kerem Oktar, Klaus Greff, Martin N Hebart, Nathan Cloos,
Nikolaus Kriegeskorte, Nori Jacoby, Qiuyi Zhang, Raja Marjieh, Robert Geirhos, Sherol Chen, Simon Kornblith, Sunayana Rane, Talia Konkle, Thomas O’Connell, Thomas Unterthiner, Andrew Kyle
Lampinen, Klaus Robert Muller, Mariya Toneva, and Thomas L. These results additionally support approximate isometry across independently-learned subject
embeddings: one rotation per subject is sufficient to map all subjects into a common space posited by
the Strong Platonic Representation Hypothesis. 4 Model–brain alignment
Finally, we test how closely our fMRI embeddings can be mapped to the intermediate representations
of artificial neural networks by fitting supervised mappings from our neural embeddings to the
final-layer embeddings of four models.Where the paper says it
Where the paper says it
[50] Radoslaw Martin Cichy, Aditya Khosla, Dimitrios Pantazis, Antonio Torralba. p. 9 Tenenbaum, Katherine M. Collins, Katherine Hermann, Kerem Oktar, Klaus Greff, Martin N Hebart, Nathan Cloos,
Nikolaus Kriegeskorte, Nori Jacoby, Qiuyi Zhang, Raja Marjieh, Robert Geirhos, Sherol Chen, Simon Kornblith, Sunayana Rane, Talia Konkle, Thomas O’Connell, Thomas Unterthiner, Andrew Kyle
Lampinen, Klaus Robert Muller, Mariya Toneva, and Thomas L.Where the paper says it
[51] Martin Schrimpf, Jonas Kubilius, Michael J. p. 9 Tenenbaum, Katherine M. Collins, Katherine Hermann, Kerem Oktar, Klaus Greff, Martin N Hebart, Nathan Cloos,
Nikolaus Kriegeskorte, Nori Jacoby, Qiuyi Zhang, Raja Marjieh, Robert Geirhos, Sherol Chen, Simon Kornblith, Sunayana Rane, Talia Konkle, Thomas O’Connell, Thomas Unterthiner, Andrew Kyle
Lampinen, Klaus Robert Muller, Mariya Toneva, and Thomas L.Where the paper says it
[52] Charlotte Caucheteux and Jean-Rémi King. Brains and algorithms partially converge in natural language processing. p. 9 Representation potentials
of foundation models for multimodal alignment: A survey. In Proceedings of the 2025 Conference on
Empirical Methods in Natural Language Processing, pages 16669–16684.Where the paper says it
[53] Talia Konkle and George A. p. 9 Tenenbaum, Katherine M. Collins, Katherine Hermann, Kerem Oktar, Klaus Greff, Martin N Hebart, Nathan Cloos,
Nikolaus Kriegeskorte, Nori Jacoby, Qiuyi Zhang, Raja Marjieh, Robert Geirhos, Sherol Chen, Simon Kornblith, Sunayana Rane, Talia Konkle, Thomas O’Connell, Thomas Unterthiner, Andrew Kyle
Lampinen, Klaus Robert Muller, Mariya Toneva, and Thomas L. We ask whether an analogous geometry can be recovered across human brains. Using fMRI data from the Natural Scenes Dataset, we propose a self-supervised
encoder that learns subject-specific embeddings from brain data alone by exploiting
repeated stimulus presentations.Where the paper says it
Where the paper says it
[54] Aria Y Wang, Kendrick Kay, Thomas Naselaris, Michael J Tarr, Leila Wehbe. p. 9 Tenenbaum, Katherine M. Collins, Katherine Hermann, Kerem Oktar, Klaus Greff, Martin N Hebart, Nathan Cloos,
Nikolaus Kriegeskorte, Nori Jacoby, Qiuyi Zhang, Raja Marjieh, Robert Geirhos, Sherol Chen, Simon Kornblith, Sunayana Rane, Talia Konkle, Thomas O’Connell, Thomas Unterthiner, Andrew Kyle
Lampinen, Klaus Robert Muller, Mariya Toneva, and Thomas L.Where the paper says it
[55] Adrien Doerig, Tim C. Kietzmann, Emily Allen.
[56] Colin Conwell, Jacob S Prince, Kendrick N Kay, George A Alvarez.
[57] Zirui Chen and Michael F. p. 9 Tenenbaum, Katherine M. Collins, Katherine Hermann, Kerem Oktar, Klaus Greff, Martin N Hebart, Nathan Cloos,
Nikolaus Kriegeskorte, Nori Jacoby, Qiuyi Zhang, Raja Marjieh, Robert Geirhos, Sherol Chen, Simon Kornblith, Sunayana Rane, Talia Konkle, Thomas O’Connell, Thomas Unterthiner, Andrew Kyle
Lampinen, Klaus Robert Muller, Mariya Toneva, and Thomas L. This suggests that current models
capture stimulus information relevant to neural responses, while still differing in the geometry of
their representational spaces. Brain-derived shared spaces may therefore provide a useful target for
identifying which dimensions of representation are shared, distorted, or missing in artificial models.Where the paper says it
Where the paper says it
[58] Matthew F. Glasser, Timothy S Coalson, Emma C Robinson, Carl D Hacker.
[59] Liang Wang, Ryan E. p. 9 Dauphin, P. Liang, and J. Synchronizing
pairwise rotations into a single shared latent space further improves cross-subject
retrieval, indicating that subject-specific spaces are mutually compatible with a
common coordinate system. These results provide evidence for a shared neural
geometry in the human visual cortex: subject-specific fMRI representations are
approximately isometric across individuals and can be translated through purely
geometric transformations.Where the paper says it
Where the paper says it
[60] Ross Wightman. Pytorch image models. p. 8 This could benefit applications such as image-to-fMRI encoding, fMRI-to-image
decoding, and synthetic neural data generation, where paired data are expensive and often unavailable
across subjects. More broadly, shared spaces provide a route for training neural models from
heterogeneous datasets in which different subjects viewed different stimuli.Where the paper says it
[61] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. p. 9 Tenenbaum, Katherine M. Collins, Katherine Hermann, Kerem Oktar, Klaus Greff, Martin N Hebart, Nathan Cloos,
Nikolaus Kriegeskorte, Nori Jacoby, Qiuyi Zhang, Raja Marjieh, Robert Geirhos, Sherol Chen, Simon Kornblith, Sunayana Rane, Talia Konkle, Thomas O’Connell, Thomas Unterthiner, Andrew Kyle
Lampinen, Klaus Robert Muller, Mariya Toneva, and Thomas L. Recent constructive work pushes this hypothesis further: if representations share a
common geometry, then embeddings from one model should be translatable into another model’s
latent space without shared inputs or paired supervision [6, 7]. However, evidence for this strong form
of representational convergence has so far come almost entirely from artificial vision and language
systems [5, 8, 9].Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Subject-specific fMRI embeddings can be translated across subjects using only intrinsic geometry via unsupervised orthogonal rotations, which
preserve distances and inner products,
allowing for cross-subject retrieval on held-out images. -
A shared latent space can be constructed by synchronizing pairwise rotations,
mapping all subject embeddings into a common coordinate system,
which improves cross-subject retrieval and indicates thatindependently learned subject spaces are mutually compatible with a common coordinate system.
-
The recovered geometry supports practical applications by suggesting that
data, encoders, and decoders learned in one subject may become transferable to another,
reducing the need forextensive subject-specific calibration
in image-to-fMRI encoding or fMRI-to-image decoding.
Abstract
The Strong Platonic Representation Hypothesis suggests that representational convergence in artificial neural networks can be harnessed constructively: embeddings can be translated across models through a universal latent space without paired data. We ask whether an analogous geometry can be recovered across human brains. Using fMRI data from the Natural Scenes Dataset, we propose a self-supervised encoder that learns subject-specific embeddings from brain data alone by exploiting repeated stimulus presentations. We show that these independently learned spaces can be translated across subjects using unsupervised orthogonal rotations, without paired cross-subject samples or intermediate model representations. Synchronizing pairwise rotations into a single shared latent space further improves cross-subject retrieval. These results provide evidence for a shared geometry in learned representations of human visual responses: subject-specific fMRI embeddings can be aligned across individuals within a common coordinate system.
Sources
- mini-vec2vec: Scaling Universal Geometry Alignment with Linear Transformations
- Representation Learning with Contrastive Predictive Coding
Related papers
- Self-Supervised Learning of Local Predictive Directions as a Neural Computation
- Synchronized Hearts, Desynchronized Brains: A Nonlinear Dynamics Perspective for Function and Failure
- Message interaction dynamics covary with brain volume across mammalian connectomes
- Planning as Dynamics Relaxation: Hippocampal Recurrent Network Realizes Optimal Goal-Directed Navigation
- The Platonic brain bridge hypothesis: brain networks as architectural priors for multimodal large language models
- BrainPFN: Amortized Brain-Behaviour Prediction via Prior-Data Fitted Networks