Platonic Representations in the Human Brain: Unsupervised Recovery of Universal Geometry

Extracted source text · arXiv 2605.20496

\documentclass{article}
\PassOptionsToPackage{numbers,sort&compress}{natbib}
% main, preprint, final

\usepackage[main, preprint]{neurips_2026}

\usepackage[utf8]{inputenc} % allow utf-8 input
\usepackage[T1]{fontenc}    % use 8-bit T1 fonts
\usepackage{hyperref}       % hyperlinks
\usepackage{url}            % simple URL typesetting
\usepackage{booktabs}       % professional-quality tables
\usepackage{amsfonts}       % blackboard math symbols
\usepackage{nicefrac}       % compact symbols for 1/2, etc.
\usepackage{microtype}      % microtypography
\usepackage{xcolor}         % colors

% Additional packages
\usepackage{amsmath}        % mathematical symbols
\usepackage{graphicx}       % figures
\usepackage{multirow}       % tables


\title{Platonic Representations in the Human Brain: Unsupervised Recovery of Universal Geometry}

\author{
  Pablo Marcos-Manchón$^{1,2}$ \quad 
  Rishi Jha$^{3}$ \quad
  Lluís Fuentemilla$^{1,2,4}$ \\
  \\
  $^{1}$Department of Cognition, Development and Education Psychology, University of Barcelona\\
  $^{2}$Institute of Neurosciences, University of Barcelona\\
  $^{3}$Department of Computer Science, Cornell University\\
  $^{4}$Bellvitge Institute for Biomedical Research, Spain
}

\begin{document}
\maketitle

\begin{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, 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.
\end{abstract}


\vspace{0.8cm} % Match spacing in the non-anonymous version

\section{Introduction}



The \emph{Platonic Representation Hypothesis} posits that independently trained artificial neural networks converge toward geometrically similar representations by recovering shared latent structure in the world~\cite{mikolov2013,Raghu2017,Kornblith2019,sucholutsky2025,Huh2024}. 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~\cite{Rishi2025,Dar2026}. However, evidence for this strong form of representational convergence has so far come almost entirely from artificial vision and language systems~\cite{Huh2024,bansal2021revisiting,lu2025survey}. Whether the same principle extends to biological neural systems remains unknown.

In neuroscience, a long line of work on inter-subject synchrony and representational similarity shows that neural responses, and the geometries they induce, can be preserved across individuals during shared stimulus processing~\cite{Hasson2004,Nastase2019,Kriegeskorte2008,Lin2024}. However, existing evidence typically depends on shared stimuli, paired measurements, or external reference spaces. Functional alignment methods for fMRI, such as \textit{hyperalignment}, map neural responses into common spaces by using shared stimuli to establish cross-subject correspondences~\cite{Haxby2011,Guntupalli2016}. Other approaches introduce anchors through model-derived feature spaces~\cite{zhou2024clipmused} or image-to-fMRI encoders~\cite{Wasserman2026}. What remains open is whether shared neural geometry can be recovered from independently learned brain representations alone.

In this paper, we extend the Strong Platonic Representation Hypothesis~\cite{Rishi2025} to the human visual cortex. We ask whether subject-specific fMRI embedding spaces, learned independently from neural data, can be translated across subjects using only the intrinsic geometry of neural responses. We evaluate this setting on the Natural Scenes Dataset (NSD)~\cite{Allen2022}, a canonical fMRI dataset of subjects viewing complex natural images. Our contributions are:
\begin{enumerate}
    \item We introduce a self-supervised encoder that learns subject-specific fMRI embeddings from repeated stimulus presentations.
    \item We show that independently learned subject embeddings are approximately isometric across brains: simple unsupervised orthogonal rotations recover accurate instance-level cross-subject correspondences.
    \item 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.
\end{enumerate}

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.

\section{Problem Setup}

Let $s \in \{1,\dots,S\}$ represent a subject who provides fMRI responses to visual stimuli drawn from a distribution $\mathcal{D}$. $X^{(s)} \in \mathbb{R}^{n_s \times v_s}$ 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 $\mathcal{Z}$ using only their intrinsic representational geometry, without paired cross-subject data for learning the mappings.

Subjects observe disjoint image sets independently sampled from $\mathcal{D}$. No shared images or paired cross-subject correspondences are available for learning the translations. A held-out set of shared images, observed by all subjects, is used only for evaluation.

\textbf{Subject-specific embeddings.}
For each subject, we learn a mapping $f_s : X^{(s)} \rightarrow Z^{(s)}$, with $Z^{(s)} \in \mathbb{R}^{n_s \times d}$, that projects voxel responses into a low-dimensional subject-specific space $\mathcal{Z}^{(s)} \subset \mathbb{R}^d$. The mapping $f_s$ is self-supervised from each subject's neural activity alone, without external model features or cross-subject supervision. Thus, $Z^{(s)}$ is intended to reflect the intrinsic organization of subject $s$'s stimulus responses.

\textbf{Unpaired brain-to-brain translation.}
Given independently learned embeddings $\{Z^{(s)}\}_{s=1}^S$, we seek transformations that translate each subject space into a common latent space. Specifically, we learn one transformation $R_s : \mathcal{Z}^{(s)} \rightarrow \mathcal{Z}$ per subject, such that embeddings evoked by the same image map to consistent coordinates, i.e., $Z^{(s)}R_s \approx Z^{(t)}R_t$ for subjects $s$ and $t$ on held-out shared images. During training, no such paired images are available; the transformations must be inferred from the global geometry of the unpaired subject-specific spaces. We evaluate translation quality using cross-subject retrieval on the held-out shared images, following prior unsupervised mapping protocols~\cite{Rishi2025,Dar2026}.

\textbf{Shared geometry hypothesis.}
We hypothesize that subject-specific spaces $\mathcal{Z}^{(s)}$ are noisy instances of a shared latent geometry. Under this hypothesis, inter-subject differences should be captured by approximately isometric transformations. We therefore restrict transformations to orthogonal maps, $R_s \in \mathcal{O}(d)$, which preserve distances and inner products. 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.

\begin{figure}
    \centering
    \includegraphics[width=\linewidth]{figures/overview.pdf}
    \caption{\textbf{Method overview.}
    \textbf{(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.
    \textbf{(B) Pairwise brain-to-brain translation.} Independently learned subject embeddings are translated between subject pairs by estimating orthogonal rotations $R_{s\rightarrow t}$ from geometry-derived pseudo-correspondences.
    \textbf{(C) Shared latent space.} Pairwise rotations are synchronized to recover one orthogonal transformation $R_s$ per subject, mapping all subject embeddings into a common space.}
    \label{fig:overview}
\end{figure}

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Method
\section{Method}\label{sec:methods}

We proceed in three stages. First, for each subject, we learn a subject-specific encoder that maps fMRI responses into a lower-dimensional embedding space. Second, we translate embeddings between subject pairs by estimating unsupervised orthogonal rotations from their geometry. Finally, we synchronize the pairwise rotations to recover one transformation per subject, mapping all embeddings into a shared latent space. An overview is shown in Fig.~\ref{fig:overview}.

\subsection{Learning geometry-preserving embeddings}

For each subject $s$, we learn a mapping $f_s$ using repeated stimulus presentations as self-supervision. Each image is presented $r$ times, yielding multiple measurements, or views, of the same underlying neural signal. For clarity, we drop the subject index $(s)$ in this subsection. Let $\{X_i\}_{i=1}^r$, with $X_i \in \mathbb{R}^{n_s \times v_s}$, denote view-specific response matrices for the same stimuli, and let $X \in \mathbb{R}^{r n_s \times v_s}$ denote their concatenation. To mitigate temporal drifts, repetitions are randomly assigned to views for each stimulus.

Traditional fMRI denoising averages repeated measurements under independence assumptions~\cite{Liu2016, Prince2022, Ozcelik2023, scotti2023reconstructing}. In contrast, we treat repetitions as noisy views of a stable latent representation and optimize $f_s$ for repetition invariance, such that for responses $x_i$ and $x_j$ to the same image, $f_s(x_i) \approx f_s(x_j)$.


\textbf{Voxel reliability weighting.}
We first reweight voxels by their reliability across repetitions. For each voxel $v \in \{1,\dots,v_s\}$, we compute reliability as the average correlation across repetition pairs:
\begin{equation}
\gamma_v = \frac{2}{r(r-1)} 
\sum_{1 \le i < j \le r}
\rho\big(X_i[:, v], \, X_j[:, v]\big).
\end{equation}

We then scale each voxel by its reliability, $\tilde{X}_i[:, v] = \gamma_v X_i[:, v]$, reducing the influence of voxels without stable repetition structure.

\textbf{Low-dimensional linear projection.}
We project reliability-weighted responses $\tilde{X}_i$ into a lower-dimensional subspace using PCA, obtaining $Y_i \in \mathbb{R}^{n_s \times d_{\text{PCA}}}$ with $d_{\text{PCA}} \ll v_s$. Let $Y \in \mathbb{R}^{r n_s \times d_{\text{PCA}}}$ denote the concatenation of all views.

To extract components shared across repetitions, we apply multi-view canonical correlation analysis (MCCA) to $\{Y_i\}_{i=1}^r$~\cite{MCCA1971, RegMCCA2011}. MCCA learns projections $\{U_i\}_{i=1}^r$, with $U_i \in \mathbb{R}^{d_{\text{PCA}} \times d}$, that maximize cross-view correlation:

\begin{equation}
\max_{\{U_i\}} \sum_{i < j} \rho\big(Y_i U_i,\; Y_j U_j\big).
\end{equation}

To obtain a single target representation, we project each sample in $Y$ through all view-specific mappings and average the projections:

\begin{equation}
\bar Z_{\mathrm{lin}} = \frac{1}{r} \sum_{i=1}^r Y U_i .
\end{equation}

We then distill these multi-view projections into a single linear mapping via ridge regression:
\begin{equation}
W^* = \arg\min_{W} \; \|Y W - \bar Z_{\mathrm{lin}}\|_F^2 + \lambda_{\mathrm{reg}} \|W\|_F^2,
\end{equation}
yielding $Z_{\mathrm{lin}} = Y W^* \in \mathbb{R}^{r n_s \times d}$.

\textbf{Nonlinear residual refinement.}
We refine the linear embedding with a residual nonlinearity:
\begin{equation}
Z = f_s(X) = Z_{\mathrm{lin}} + \alpha \, g_\theta(Z_{\mathrm{lin}}),
\end{equation}
where $g_\theta$ is a multi-layer perceptron (MLP) and $\alpha$ is a learnable scalar. We freeze the linear projection and optimize only $\theta$ and $\alpha$. Given view embeddings $\{Z_i\}_{i=1}^r$, where $Z_i=f_s(X_i)$, we use a contrastive InfoNCE loss over all view pairs:

\begin{equation}
\mathcal{L}_{\mathrm{NCE}} = \frac{1}{r(r-1)} \sum_{\substack{i,j=1 \\ i \neq j}}^{r} \mathcal{L}_{\mathrm{InfoNCE}}(Z_i, Z_j),
\end{equation}
with in-batch negatives and cosine similarity~\cite{infoNCE2019}. We also add a cosine pull term,

\begin{equation}
\mathcal{L}_{\mathrm{pull}} = \frac{2}{r(r-1)} \sum_{1 \le i < j \le r}
\big(1 - \mathrm{sim}(Z_i, Z_j)\big),
\end{equation}
where $\mathrm{sim}(Z_i, Z_j)$ is the mean cosine similarity between corresponding samples. The refinement minimizes $(\theta^*, \alpha^*) = \arg\min_{\theta,\alpha} \mathcal{L}_{\mathrm{NCE}} + \lambda_{\mathrm{pull}}\mathcal{L}_{\mathrm{pull}}$.

\subsection{Pairwise brain-to-brain translation}\label{subsec:pairwise}

To translate embeddings between two subjects, we adapt \textit{mini-vec2vec}~\cite{Dar2026} to neural embeddings. Given subjects $s$ and $t$, we learn an orthogonal transformation $R_{s \rightarrow t} \in \mathcal{O}(d)$ such that $Z^{(s)}R_{s \rightarrow t} \approx Z^{(t)}$, without paired cross-subject samples during training.

We average repetitions in embedding space to obtain one representation per image, $Z^{(s)} \in \mathbb{R}^{n_s \times d}$ and $Z^{(t)} \in \mathbb{R}^{n_t \times d}$, reducing trial-level measurement error. We construct pseudo-matched pairs by clustering each space with K-means and matching centroids through their pairwise similarity structure using a quadratic assignment solver. Each embedding in $Z^{(s)}$ is then matched to the average of its nearest neighbors in $Z^{(t)}$ based on relative similarity to these matched anchors, yielding pseudo-parallel pairs $(z^{(s)}, \tilde z^{(t)})$. These pairs define the initial orthogonal Procrustes problem:

\begin{equation}
R_{s \rightarrow t}^{(0)} = \arg\min_{R \in \mathcal{O}(d)} \| Z^{(s)} R - \tilde Z^{(t)} \|_F^2.
\end{equation}

We refine the translation iteratively using an approach similar to Iterative Closest Point~\cite{ICP}. At iteration $k$, a subset of source embeddings is transformed by $R_{s \rightarrow t}^{(k)}$ and matched to nearest neighbors in the target space. These updated pseudo-targets define a new Procrustes solution $\widehat R_{s \rightarrow t}^{(k)}$, and the transformation is updated as

\begin{equation}
\tilde R_{s \rightarrow t}^{(k+1)} = (1-\beta)R_{s \rightarrow t}^{(k)} + \beta \widehat R_{s \rightarrow t}^{(k)}.
\end{equation}

After each update, $\tilde R_{s \rightarrow t}^{(k+1)}$ is projected onto $\mathcal{O}(d)$ by SVD to obtain $R_{s \rightarrow t}^{(k+1)}$.

Finally, we symmetrize the pairwise translations:

\begin{equation}
R_{s \rightarrow t} = \mathrm{Proj}_{\mathcal{O}(d)}\!\left( \frac{R_{s \rightarrow t} + R_{t \rightarrow s}^\top}{2} \right).
\end{equation}

This enforces $R_{t \rightarrow s} = R_{s \rightarrow t}^\top$, which is required for global synchronization and improves stability. Details on random-seed selection and rotation stability are provided in the Appendix \ref{sec:details}.

\subsection{Shared latent space construction}\label{subsec:global}

Given pairwise translations $\{R_{s \rightarrow t}\}_{s,t=1}^S$, we construct a shared latent space $\mathcal{Z}$ by solving an orthogonal synchronization problem over $\mathcal{O}(d)$~\cite{Singer2011,WangRotation2013}. The goal is to recover one transformation $R_s \in \mathcal{O}(d)$ per subject such that $R_{s \rightarrow t} \approx R_s R_t^\top$, placing all subjects in a common coordinate system, i.e., $Z^{(s)}R_s \approx Z^{(t)}R_t$.

We form a block matrix $B \in \mathbb{R}^{Sd \times Sd}$ whose $(s,t)$-th block is $R_{s \rightarrow t}$ for $s \neq t$ and the identity for $s=t$. In the ideal noise-free case, this matrix factorizes as
\begin{equation}
B = 
\begin{bmatrix}
R_1 \\
R_2 \\
\vdots \\
R_S
\end{bmatrix}
\begin{bmatrix}
R_1 \\
R_2 \\
\vdots \\
R_S
\end{bmatrix}^\top.
\end{equation}

We recover a relaxed solution using a spectral method for orthogonal synchronization~\cite{Singer2011}. We compute the top-$d$ eigenvectors of $B$, yielding $U \in \mathbb{R}^{Sd \times d}$, whose blocks $\{U_s\}_{s=1}^S$ approximate the subject-specific transformations. Each block $U_s$ is projected onto $\mathcal{O}(d)$ by taking its closest orthogonal matrix.

The resulting transformations define a shared latent space in which each subject embedding is mapped as $Z_{\mathrm{shared}}^{(s)} = Z^{(s)}R_s$. Global synchronization denoises pairwise estimates by enforcing cycle consistency across the subject graph.

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Experiments
\section{Experiments}

We evaluate our method on the Natural Scenes Dataset (NSD)~\cite{Allen2022}, an fMRI dataset of 8 participants viewing natural images from COCO~\cite{Lin2014MicrosoftCC}. Each participant viewed up to 10,000 distinct images, each repeated up to three times. NSD includes subject-specific images, unique to each participant, and a smaller set of images shared across participants. We use only subject-specific, non-shared images to learn subject encoders and brain-to-brain translations; shared images are held out exclusively for evaluation. We restrict evaluation to shared images with three repetitions for every subject, yielding 515 images. See Appendix~\ref{sec:fmri-processing} for preprocessing details.

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. Across subjects, repetitions are first averaged, and translated embeddings from subject $s$ retrieve the matching image from subject $t$ among the same 515 held-out images. In both settings, we report Mean Rank, R@1, and RSA. Mean Rank is the average rank of the correct image (chance $=258$, optimum $=1$); R@1 is nearest-neighbor accuracy (chance $=1/515 \approx 0.002$); RSA is the Pearson correlation between representational dissimilarity matrices.

\begin{table}[b]
\centering
\caption{\textbf{Per-subject encoder performance.}
Within-subject retrieval across repeated presentations, averaged across repetition pairs. Embeddings from one repetition are used to retrieve the matching image from another among 515 candidates. RSA is the Pearson correlation between RDMs across repetitions. Lower is better for Mean Rank (chance $=258$); higher is better for R@1 (chance $=0.002$) and RSA.}
\label{tab:within_subject_per_subject}
\resizebox{\linewidth}{!}{%
    \input{tables/table_within_subject.tex}
}
\end{table}

\subsection{Within-subject encoder evaluation}

We first evaluate whether the subject-specific encoder extracts stable stimulus representations from noisy repeated fMRI responses, using the within-subject retrieval protocol defined above. Encoder ablations and hyperparameter configurations are provided in Appendix~\ref{sec:ablations}.



\textbf{Performance across subjects.}
Table~\ref{tab:within_subject_per_subject} reports full-encoder performance for each subject. In general our embeddings are stable: with an average rank of $5.28$, within-subject retrieval works well across subjects. S1 and S2 are nearly perfectly matched across repetitions (Mean Rank $\approx 1$; R@1 $>0.98$), whereas S3 and S8 show lower, but still solid numbers. 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.

\textbf{Comparison with encoder baselines.}
Table~\ref{tab:within_subject_main} compares our encoder against three baseline families. First, we include direct neural baselines using preprocessed GLMsingle~\cite{GLMSingle2022} responses and PCA-reduced fMRI responses. Second, we compare against multiview methods trained self-supervised from repeated presentations. Third, we include model-guided baselines, where fMRI responses are linearly regressed to pretrained model embeddings. Direct neural baselines provide limited retrieval, while multiview methods improve instance-level matching but yield low RSA. Model-guided baselines preserve stronger geometry, but are weaker at retrieving repeated presentations of the same image. Our full encoder achieves the lowest Mean Rank and highest R@1, with RSA comparable to the strongest model-guided baselines. Thus, repetition-based self-supervision recovers stronger instance-level image information than externally guided encoders, likely because it directly optimizes invariance across neural measurements of the same stimulus rather than fitting an intermediate model space (see Appendix~\ref{sec:details} for detailed per-method results and details).

\begin{table}[h]
\centering
\caption{\textbf{Within-subject encoder comparison.}
Retrieval and RSA across repeated presentations, averaged across subjects and repetition pairs. Mean Rank and R@1 measure instance-level stability; RSA measures geometry preservation across repetitions. Lower is better for Mean Rank (chance $=258$); higher is better for R@1 (chance $=0.002$) and RSA.}
\label{tab:within_subject_main}
\resizebox{\linewidth}{!}{%
\input{tables/table_within_subject_main.tex}
}
\end{table}

\subsection{Pairwise brain-to-brain translation}

Next, we test whether fMRI embeddings from one subject can be translated into another subject's space using only subject-specific encoders and an unsupervised orthogonal map. Each encoder is trained independently on subject-specific, non-overlapping images. To reduce trial-level measurement error, embeddings are averaged across repetitions before translation, yielding one representation per image per subject (Subsection~\ref{subsec:pairwise}). For each ordered pair $(s,t)$, we learn an orthogonal transformation $R_{s \rightarrow t}$ using only non-shared images. We evaluate on the 515 held-out shared images using cross-subject retrieval: each translated embedding from subject $s$ is used to retrieve the matching image among all embeddings from subject $t$. Because unsupervised translation is sensitive to initialization, we follow prior unsupervised mapping protocols~\cite{Rishi2025,Dar2026} and run the pairwise translation procedure with 10 random seeds for each ordered subject pair, reporting the best-performing seed in the main results. Appendix~\ref{sec:rotation_stability} reports seed-averaged rotations and stability across seeds.

Fig.~\ref{fig:pairwise_alignment} shows that our method achieves low rank, high recall, and high RSA across most subject pairs. Our method outperforms all baselines on retrieval metrics (Table~\ref{tab:pairwise_alignment}) and performs comparably to external-reference baselines on RSA, despite not using pretrained model spaces to learn the translations. No-translation controls perform poorly, confirming that subject embeddings are not directly comparable in their native coordinate systems. Optimal-transport matching also performs poorly, suggesting that learning orthogonal translations is critical for recovering cross-subject correspondences.

\begin{figure}[t]
    \centering
    \includegraphics[width=\linewidth]{figures/pairwise_alignment_results.pdf}
    \caption{\textbf{Pairwise brain-to-brain translation.}
Performance for each ordered subject pair $(s,t)$. Embeddings from source subject $s$ are mapped into target subject $t$'s space using the unsupervised orthogonal transformation $R_{s \rightarrow t}$ and evaluated on the 515 held-out shared images. Mean Rank (average $\pm$ std: $2.56 \pm 1.71$, chance $=258$) and R@1 (average $\pm$ std: $0.78 \pm 0.14$, chance $=0.002$) measure image-level retrieval after translation. RSA is the Pearson correlation between subject-specific RDMs. Darker colors indicate better performance.}
    \label{fig:pairwise_alignment}
\end{figure}


\begin{table}[ht]
\centering
\caption{\textbf{Pairwise brain-to-brain translation baselines.}
Performance is reported as the average across ordered off-diagonal subject pairs on the 515 held-out shared images. Rows compare no-translation controls, optimal-transport matching, external model-space references, and our unpaired orthogonal translation. Lower is better for Mean Rank (chance $=258$); higher is better for R@1 (chance $=0.002$) and RSA.}
\label{tab:pairwise_alignment}
\resizebox{\linewidth}{!}{%
    \input{tables/table_pairwise_alignment.tex}
}
\end{table}

\subsection{A Platonic brain-to-brain translation layer}

Pairwise translations map individual subject pairs, but do not guarantee a single coherent coordinate system across subjects. We therefore integrate the pairwise rotations into a all-subject shared latent space by recovering one orthogonal transformation $R_s$ per subject through global synchronization (Subsection~\ref{subsec:global}), mapping all embeddings into a common coordinate system (Fig.~\ref{fig:overview}C).

Fig.~\ref{fig:global_alignment} shows that the synchronized space not only supports accurate retrieval across subject pairs, but also, when compared with independent pairwise translations, improves average Mean Rank from $2.47$ to $1.97$ and R@1 from $0.79$ to $0.83$ across ordered off-diagonal pairs. This indicates that enforcing consistency across the subject graph denoises pairwise estimates and yields a coherent shared latent space. 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.

\begin{figure}
    \centering
    \includegraphics[width=\linewidth]{figures/pairwise_alignment_global_results.pdf}
    \caption{\textbf{Shared-space brain-to-brain translation.}
    Retrieval performance after synchronizing pairwise brain-to-brain rotations into a single shared latent space. Each subject is mapped into the common coordinate system using one orthogonal transformation $R_s$, and retrieval is evaluated across ordered subject pairs on the 515 held-out shared images. Left: Mean Rank, lower is better (average $\pm$ std: $2.00 \pm 0.76$; chance $=258$). Right: R@1, higher is better (average $\pm$ std: $0.83 \pm 0.09$; chance $=0.002$). Diagonal entries are omitted. Darker colors indicate better performance in both panels.}
    \label{fig:global_alignment}
\end{figure}

\subsection{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. We use our encoder training set to fit the translation, and evaluate retrieval on the held-out shared images. For translators, we compare a semi-orthogonal map which combines dimensionality matching with an orthogonality constraint and ridge regression.

Table~\ref{tab:model_brain_alignment} shows that semi-orthogonal model--brain mappings recover moderate instance-level correspondence (best Mean Rank $=13.33$, R@1 $=0.29$), but remain weaker than brain-to-brain translations. Ridge regression improves retrieval and RSA (best Mean Rank $=5.86$, R@1 $=0.51$), indicating that model features are predictive of the recovered neural embeddings when more flexible linear transformations are allowed. However, these supervised model--brain mappings remain below the shared-space brain-to-brain translation results (Mean Rank $=1.97$, R@1 $=0.83$), suggesting that the tested model spaces overlap with neural embeddings in stimulus information but are not related to them by the same near-isometric transformations observed across subjects.

\begin{table}[t]
\centering
\caption{\textbf{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. We compare semi-orthogonal maps, which preserve geometry after dimensionality matching, with ridge regression, which allows a more flexible linear transformation. Values are averaged across subjects. Lower is better for Mean Rank; higher is better for R@1 and RSA.}
\label{tab:model_brain_alignment}
\resizebox{0.9\linewidth}{!}{%
    \input{tables/table_model_brain_alignment}
}
\end{table}



%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Related work
\section{Related work}

\textbf{Universal geometry in machine learning.}
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~\cite{Li2016,mikolov2013,Raghu2017,Kornblith2019}. These observations motivate the \textit{Platonic Representation Hypothesis}~\cite{Huh2024}, which proposes that models converge toward a shared latent structure as scale and data coverage increase. This convergence has been studied through representational similarity measures~\cite{Kornblith2019} and constructive methods such as model stitching and latent-space mapping~\cite{bansal2021revisiting,AlvarezMelis2018,Grave2019,Moschella2022}. Recent work tests a stronger version of this hypothesis by showing that embedding spaces can be translated from intrinsic geometry alone, without shared inputs or paired supervision~\cite{Rishi2025,lample2018word,Dar2026}. In this work, we ask whether this constructive form---the Strong Platonic Representation Hypothesis---also holds for biological neural representations.

\textbf{Shared representations in neuroscience.}
Shared neural representations are harder to recover because brain data are noisy, limited, and subject to anatomical and functional variability. Nevertheless, inter-subject synchrony~\cite{Hasson2004,Nastase2019} and representational similarity analyses~\cite{Kriegeskorte2008,Lin2024} show that neural response structure is partly preserved across subjects~\cite{MarcosManchon2025,Chen2017,Diedrichsen2017}. Functional alignment methods such as hyperalignment map subjects into common spaces, but typically require shared stimuli to establish correspondences~\cite{Haxby2011,Guntupalli2016}. Other methods introduce anchors through model-derived feature spaces~\cite{zhou2024clipmused} or image-to-fMRI encoders~\cite{Wasserman2026}, while unsupervised alternatives infer correspondences through structural or distributional matching~\cite{Nakamura2024} and optimal transport-based functional alignment~\cite{Bazeille2019,Bazeille2021}. 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.

\textbf{Model--brain alignment.}
A complementary literature aligns neural activity with deep neural network representations, revealing systematic correspondences between cortical processing stages and model layers~\cite{Yamins2014,Khaligh2014,Cichy2016,MarcosManchon2025}. These approaches support encoding, decoding, and benchmarking efforts such as \emph{Brain-Score}~\cite{Schrimpf2020,Caucheteux2022,Konkle2022,Wang2023,Doerig2025,Conwell2024}. Recent work further suggests that brains and models may share low-dimensional representational axes~\cite{Chen2025}. However, model-mediated approaches impose or evaluate neural representations through an external feature geometry. Our method instead recovers a shared brain space directly from neural data, using model--brain mappings only diagnostically.

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Discussion
\section{Discussion}

The Strong Platonic Representation Hypothesis \cite{Rishi2025} 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. Whereas the original hypothesis was conjectured over artificial embeddings, in this work, we evaluated the prediction on the human brain. We found that subject-specific embeddings learned from brain data alone can be translated across individuals using unsupervised orthogonal rotations, without paired cross-subject samples or intermediate model representations. 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.

This shared neural geometry has practical implications for cross-subject neural modeling. If subject-specific 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 subject-specific calibration. 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.

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. 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.

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. It remains unclear whether similar isometries will hold in lower-SNR modalities, smaller datasets, or higher-level cognitive tasks. The unpaired translation procedure is also sensitive to initialization, requiring repeated runs and seed selection; developing fully internal model-selection criteria is an important next step. 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.





%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Acknowledgments
\begin{ack}
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). We thank CERCA Programme/Generalitat de Catalunya for institutional support. R.J. was supported by the Digital Life Initiative Doctoral Fellowship at Cornell Tech.
\end{ack}

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% References
{
\small
\bibliographystyle{unsrtnat}
\bibliography{references}
}

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

\newpage
\appendix
\input{supplementary}


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

%\newpage
%\input{checklist.tex}

\end{document}
\begin{tabular}{ll|cccccccc|c}
\toprule
\textbf{Method} & \textbf{Metric} & \textbf{S1} & \textbf{S2} & \textbf{S3} & \textbf{S4} & \textbf{S5} & \textbf{S6} & \textbf{S7} & \textbf{S8} & \textbf{Avg} \\
\midrule
\multirow{3}{*}{\textbf{Ours (full)}} 
& Mean Rank $\downarrow$ & 1.03 & 1.12 & 11.25 & 4.69 & 1.77 & 3.09 & 7.6 & 11.67 & 5.28 \\
& R@1 $\uparrow$         & 0.99 & 0.98 & 0.75 & 0.79 & 0.87 & 0.83 & 0.72 & 0.62 & 0.82 \\
& RSA $\uparrow$         & 0.60 & 0.61 & 0.47 & 0.51 & 0.60 & 0.52 & 0.5 & 0.43 & 0.53 \\
\bottomrule
\end{tabular}
\begin{tabular}{ll|ccc}
\toprule
\textbf{Method} & \textbf{Setting} & \textbf{Mean Rank} $\downarrow$ & \textbf{R@1} $\uparrow$ & \textbf{RSA} $\uparrow$ \\
\midrule
Preprocessed fMRI (GLMsingle \cite{GLMSingle2022}) & GLM (repetitions) & 144.15 & 0.07 & 0.18 \\
\midrule
PCA & Unsupervised & 78.1 & 0.19 & 0.27 \\
PCA + SplitAE \cite{SplitAE} & SSL Repetitions & 93.78 & 0.07 & 0.29 \\
% PCA + GCCA & Repetitions & - & - & - \\
% PCA + CCA \cite{RegMCCA2011} & SSL Repetitions & 23.23 & 0.52 & 0.19 \\
PCA + DeepCCA \cite{DeepCCA} & SSL Repetitions & 23.21 & 0.52 & 0.19 \\
PCA + Multiview CCA \cite{RegMCCA2011} & SSL Repetitions & 19.23 & 0.60 & 0.21 \\
% Rel. weighting + PCA + DeepCCA & Repetitions & 14.52 & 0.67 & 0.21 \\
\midrule
DINOv2 ViT-S/14 \cite{dinov2} + Ridge & Vision Model Guided & 24.82	 & 0.30	 & 0.46  \\ % vit_small_patch14_dinov2
CLIP ViT-B/16 \cite{Radford2021} + Ridge & Vision Model Guided & 23.23 & 0.26 & 0.50  \\ % vit_base_patch16_clip_224
ViT AugReg-L/16 \cite{steiner2022how} + Ridge & Vision Model Guided & 20.3 & 0.32 & 0.53 \\ % vit_large_patch16_224.augreg_in21k_pool
all-MiniLM-L6-v2 \cite{SentenceBert} + Ridge & Language Model Guided & 30.47 & 0.18 & \textbf{0.56} \\ % nsd_language_embeddings_all-mini
%\midrule
%Mind-Vis (VAE) & Generative &  &  &  \\
\midrule
% Ours (linear) & SSL Repetitions & 11.73 & 0.73 & 0.23 \\
Ours (full) & SSL Repetitions & \textbf{5.28} & \textbf{0.82} & 0.53 \\
\bottomrule
\end{tabular}
\begin{tabular}{ll|ccc}
\toprule
\textbf{Representation} & \textbf{Alignment}
& \textbf{Mean Rank} $\downarrow$ & \textbf{R@1} $\uparrow$ & \textbf{RSA} $\uparrow$ \\
\midrule

%\multicolumn{5}{l}{\textit{No cross-subject alignment}} \\
fMRI betas + PCA & No alignment & 189.28 & 0.007 & 0.39 \\

\midrule
%\multicolumn{5}{l}{\textit{Optimal transport matching (transductive / oracle baseline)}} \\
fMRI betas + PCA & Entropic GW on test set & 166.04 & 0.011  & 0.39 \\
Ours embeddings (full) & Entropic GW on test set & 80.96 & 0.465  & 0.64 \\

\midrule
%\multicolumn{5}{l}{\textit{External model-defined shared spaces}} \\
ViT AugReg-L/16~\cite{steiner2022how} + Ridge & Vision Model Guided & 5.54 & 0.48 & 0.73 \\
CLIP ViT-B/16~\cite{Radford2021} + Ridge & Vision Model Guided & 6.60 & 0.44 & 0.70 \\
DINOv2 ViT-S/14~\cite{dinov2} + Ridge & Vision Model Guided & 7.00 & 0.47 & 0.67 \\
all-MiniLM-L6-v2~\cite{SentenceBert} + Ridge & Language Model Guided & 7.93 & 0.35 & \textbf{0.76} \\

\midrule
%\multicolumn{5}{l}{\small{\textit{Unsupervised alignment with unpaired images}}} \\
Ours embeddings (full) & Unsupervised orthogonal & \textbf{2.56} & \textbf{0.78} & 0.63 \\

\bottomrule
\end{tabular}
\begin{tabular}{lccc|ccc}
\toprule
\multirow{2}{*}{\textbf{Source model}} 
& \multicolumn{3}{c|}{\textbf{Semi-orthogonal mapping}} 
& \multicolumn{3}{c}{\textbf{Ridge regression}} \\
\cmidrule(lr){2-4} \cmidrule(lr){5-7}
& \textbf{Mean Rank} $\downarrow$ & \textbf{R@1} $\uparrow$ & \textbf{RSA} $\uparrow$
& \textbf{Mean Rank} $\downarrow$ & \textbf{R@1} $\uparrow$ & \textbf{RSA} $\uparrow$ \\
\midrule
all-MiniLM-L6-v2 & 35.62 & 0.08 & 0.31 & 27.05 & 0.11 & 0.45 \\
CLIP ViT-B/16 & 21.65 & 0.23 & 0.23 & 7.86 & 0.44 & 0.58 \\
ViT AugReg-L/16 & \textbf{13.33} & \textbf{0.29} & \textbf{0.41} & 8.85 & 0.41 & 0.57 \\
DINOv2 ViT-S/14 & 22.71 & 0.24 & 0.24 & \textbf{5.86} & \textbf{0.51} & \textbf{0.57} \\
\bottomrule
\end{tabular}
% Format the numbering of the figures, tables and eqs.
\setcounter{figure}{0}
\setcounter{table}{0}
\setcounter{equation}{0}
\renewcommand{\thefigure}{S\arabic{figure}}
\renewcommand{\thetable}{S\arabic{table}}
\renewcommand{\theequation}{S\arabic{equation}}

\begin{center}
\Large{\textbf{Appendix}}
\end{center}

\section{fMRI preprocessing details}\label{sec:fmri-processing}

For all main analyses, we use the Natural Scenes Dataset (NSD)~\cite{Allen2022} volumetric 1\,mm GLM beta estimates\cite{GLMSingle2022}, version 3, as provided by the dataset authors\footnote{\url{https://registry.opendata.aws/nsd/} (accessed May 2026).}. For each subject, we extract voxels from the \texttt{nsdgeneral} ROI, a reliability-based mask that includes visually responsive voxels across visual and high-level associative cortex. This yields approximately $\sim$85K voxels per subject. The resulting responses are organized into a trial-by-voxel matrix $X \in \mathbb{R}^{n_{\mathrm{trials}} \times v_s}$ for each subject $s$.
 
Before fitting the encoder, we apply a simple normalization procedure to reduce the effect of extreme beta values. For each subject, voxel values are clipped at the 0.05th and 99.95th percentiles and stored as the input response matrix for subsequent preprocessing. We additionally remove acquisition-related confounds by fitting a linear model of session/run structure on the training trials and subtracting the predicted confound component from all responses. All train/test splits are defined at the stimulus level before fitting any preprocessing transformations.


\subsection{Confound removal}
NSD data were acquired over multiple scanning sessions, with each subject completing up to 40 sessions and multiple runs per session. We observed that trial-by-trial correlation matrices contained visible block structure aligned with run/session order, indicating acquisition-related components in the raw fMRI responses (Fig.~S1A--B). To reduce this structure, we residualized each subject's voxel responses with respect to run/session confounds.

Let $X \in \mathbb{R}^{n_{\mathrm{trials}} \times v_s}$ denote the preprocessed beta matrix and let $C \in \mathbb{R}^{n_{\mathrm{trials}} \times c}$ denote a confound design matrix encoding the run/session membership of each trial. We fit a ridge-regularized linear confound model using only training trials:
\[
B^* = \arg\min_B \|X_{\mathrm{train}} - C_{\mathrm{train}}B\|_F^2
+ \lambda_{\mathrm{conf}}\|B\|_F^2 .
\]
The fitted confound coefficients are then applied to all trials, and residualized responses are computed as
\[
X_{\mathrm{res}} = X - CB^* .
\]
Thus, nuisance parameters are estimated exclusively from the training set and then applied to both training and held-out trials, avoiding leakage from the evaluation responses into the fitted residualization model.

\begin{figure}[!b]
    \centering
    \includegraphics[width=\linewidth]{figures/residualization_effects.pdf}
    \caption{\textbf{Residualization mitigates acquisition-related trial structure.}
    Trial-by-trial correlation matrices for the first 250 trials from the first session of subject S1, shown for raw fMRI responses and linear encoder embeddings before and after residualizing run/session confounds. Trials are ordered by acquisition order; therefore, block-like off-diagonal structure reflects correlations due to scanning confounds rather than stimulus identity. Raw fMRI responses show pronounced run/session structure before residualization, which is partially mitigated after confound regression. The linear encoder further attenuates acquisition-related correlations. Diagonal entries are omitted for visualization.}
    \label{fig:residual_correlations}
\end{figure}

Residualization partially mitigates the run/session block structure visible in trial-wise correlation matrices (Fig.~\ref{fig:residual_correlations}). All experiments and baselines in the main text use these residualized fMRI responses as input.



\subsection{Brain region selection}
For the main analyses, we extract neural responses from the NSDGeneral ROI provided by the NSD authors~\cite{Allen2022}. This mask includes visually responsive voxels across visual and high-level associative cortex selected based on stimulus-related reliability. We use the volumetric preprocessing for the main experiments.

Table~\ref{tab:roi_preprocessing_ablation} evaluates the effect of cortical region and preprocessing space using the same encoder configuration as in the main analyses (see Section~\ref{sec:ablations}). Encoder quality is measured with within-subject retrieval across repeated presentations. Performance improves as the cortical mask expands from V1 to V1--V4 and NSDGeneral, with NSDGeneral yielding the best retrieval and RSA. Volumetric and surface-based NSDGeneral preprocessing give similar performance, so we use volumetric NSDGeneral responses for all main reported analyses.

\begin{table}[ht]
\centering
\caption{\textbf{Effect of brain region and preprocessing on encoder performance.}
Within-subject retrieval across repeated presentations for different cortical regions (V1, V1--V4, NSDGeneral, Kastner, and visual streams; Left and Right hemispheres) and preprocessing spaces (volumetric and surface-based). Results are averaged across subjects and repetition pairs. Lower is better for Mean Rank (chance $=258$); higher is better for R@1 (chance=0.002) and RSA.}
\label{tab:roi_preprocessing_ablation}
\resizebox{\linewidth}{!}{%
    \input{tables/table_ablation_rois}
}
\end{table}

\section{Encoder details and ablations}\label{sec:ablations}

\subsection{Encoder implementation details}

All subject encoders use the same hyperparameters. For the linear stage, we use $d_{\mathrm{PCA}}=768$ PCA components and an embedding dimensionality of $d=128$. The nonlinear residual refinement is a one-hidden-layer MLP with hidden size $d_{\mathrm{hidden}}=768$. The residual scaling parameter $\alpha$ is learned jointly with the MLP parameters and converges to $\alpha \approx 0.4$ across subjects.

The MLP is optimized with Adam for 2000 training steps using the combined objective described in the main text, with $\lambda_{\mathrm{NCE}}=1$ and $\lambda_{\mathrm{pull}}=0.5$. Hyperparameters were selected by within-subject retrieval across repeated presentations, balancing Mean Rank and R@1. Unless stated otherwise, all experiments use these settings, yielding 128-dimensional subject embeddings.

\begin{table}[b]
\centering
\caption{\textbf{Encoder component ablation.}
Within-subject retrieval across repeated presentations after removing components of the encoder. Results are averaged across subjects and repetition pairs. Lower is better for Mean Rank; higher is better for R@1 and RSA.}
\label{tab:encoder_component_ablation}
\resizebox{\linewidth}{!}{%
    \input{tables/table_ablation_encoder_stages}
}
\end{table}


\subsection{Encoder component ablation.}
We ablated the main components of the encoder by removing reliability weighting, MCCA, and the nonlinear residual refinement. Table~\ref{tab:encoder_component_ablation} shows that the full encoder gives the best retrieval performance. The nonlinear refinement contributes strongly to RSA and retrieval, while the full combination of reliability weighting, PCA, MCCA, and nonlinear refinement yields the best Mean Rank and R@1.

\subsection{Encoder dimensionality sensitivity.}
We evaluated within-subject retrieval across a hyperparameter grid varying PCA dimensionality $d_{\mathrm{PCA}} \in \{512,768,1024,1280\}$, embedding dimensionality $d \in [64,384]$, hidden size $d_{\mathrm{hidden}} \in \{128,256,512,768,1024\}$, and MLP depth $\in \{1,2,3\}$. Fig.~\ref{fig:encoder_dimensionality_sensitivity} summarizes the search by plotting Mean Rank as a function of $d$, with separate curves for $d_{\mathrm{PCA}}$ and panels for MLP depth. Since $d_{\mathrm{hidden}}$ is not shown, each point reports the best Mean Rank across hidden sizes for that combination of $d$, $d_{\mathrm{PCA}}$, and depth. Overall, a single hidden layer performs best, and the strongest stable configurations use $d_{\mathrm{PCA}}=768$ or $1024$ with embedding dimensionality around $d=100$--$150$.

\begin{figure}[h]
    \centering
    \includegraphics[width=\linewidth]{figures/pairwise_alignment_ablations_dimensions.pdf}
     \caption{\textbf{Encoder dimensionality sensitivity.}
    Mean Rank across encoder dimensionalities. Lower is better. Each point reports the best value across $d_{hidden}$ sizes. Average mean rank across subjects.}
    \label{fig:encoder_dimensionality_sensitivity}
\end{figure}

\section{Experimental details}\label{sec:details}

This section reports additional details to reproduce the evaluation and main results. 



\subsection{Within-subject encoder baseline details}

To assess whether encoder improvements were consistent across subjects, Table~\ref{tab:within_subject_baselines_subjects} reports Mean Rank for each within-subject baseline separately for all subjects (extending Table \ref{tab:within_subject_main}).

\begin{table}[h]
\centering
\caption{\textbf{Per-subject within-subject retrieval across encoder baselines.}
Mean Rank for within-subject retrieval across repeated presentations, reported separately for each subject and averaged across repetition pairs. Lower is better; chance Mean Rank is $258$.}
\label{tab:within_subject_baselines_subjects}
\resizebox{\linewidth}{!}{%
    \input{tables/table_detailed_within_subject}
}
\end{table}

We evaluated model-guided baselines by fitting a linear ridge regression from PCA-reduced fMRI responses (768 components) to intermediate model representations, and then measuring within-subject retrieval across repeated presentations. Vision embeddings were extracted from the last layer of pretrained models applied to the NSD images using the TIMM library\cite{rw2019timm}, while language embeddings were extracted from the COCO captions associated with each image using sentence-transformers\cite{SentenceBert}. Language models follow those used in~\cite{Dar2026}, and vision models were selected from~\cite{MarcosManchon2025}. A complete list of model identifiers is provided as a Hugging Face collection\footnote{https://huggingface.co/collections/pablomm/platonic-representations-in-the-human-brain}. The main text reports representative models from the main model families in Table~\ref{tab:within_subject_main}. Table~\ref{tab:model_guided_baselines_subjects} provides the detailed per-subject evaluation for the larger set of model representations tested.

\begin{table}[h]
\centering
\caption{\textbf{Per-subject model-guided ridge baselines.}
Mean Rank for within-subject retrieval across repeated presentations using ridge mappings from fMRI responses to pretrained vision-model embeddings. Results are reported separately for each subject and averaged across subjects. Lower is better; chance Mean Rank is $258$.}
\label{tab:model_guided_baselines_subjects}
\resizebox{\linewidth}{!}{%
    \input{tables/table_within_subject_models}
}
\end{table}

\subsection{Pairwise rotation selection}\label{sec:rotation_stability}

The pairwise brain-to-brain translation procedure is sensitive to initialization because pseudo-correspondences are obtained from unsupervised centroid matching and iterative nearest-neighbor refinement, as also reported in the original mini-vec2vec setting for language models~\cite{Dar2026}. This sensitivity is amplified in the neural setting, where the number of unpaired training samples is limited and embeddings contain measurement noise. Following prior unsupervised mapping protocols~\cite{Rishi2025,Dar2026}, we therefore run the translation procedure multiple times with different random seeds for each ordered subject pair and report the best-performing candidate.

For the main experiments, we run each ordered subject pair with 10 random seeds. Fig.~\ref{fig:rotation-stability} reports the best-performing seed for each pair, the mean rotation obtained by averaging the 10 rotations and projecting the result back onto $\mathcal{O}(d)$, and the distribution of candidate performance across all seeds and subject pairs.



\begin{figure}[ht]
    \centering
    \includegraphics[width=\linewidth]{figures/best_mean_candidate_vs_mean_candidate.pdf}
    \caption{\textbf{Pairwise rotation stability across random seeds.}
    For each ordered subject pair, we ran the unsupervised pairwise brain-to-brain translation procedure with 10 different random seeds and evaluated Mean Rank on the cross-subject retrieval task using the 515 held-out shared images. Left: Mean Rank for the best-performing seed for each pair, corresponding to the protocol used in the main results (average $\pm$ std: $2.56 \pm 1.71$). Middle: Mean Rank after averaging the 10 rotation matrices for each pair and projecting the result back onto $\mathcal{O}(d)$, using all runs without seed selection (average $\pm$ std: $4.61 \pm 4.37$). Right: distribution of Mean Rank values across all candidate rotations and subject pairs ($10$ seeds $\times$ $56$ ordered pairs). Lower is better; darker colors in the heatmaps indicate better retrieval.}
    \label{fig:rotation-stability}
\end{figure}


\subsection{Compute requirements}
All experiments were run on a single workstation with an NVIDIA A40 GPU (48 GB) and an AMD Ryzen 9 CPU. The experiments are not computationally intensive: training the linear encoder for one subject takes approximately 2 minutes on CPU, training the nonlinear residual MLP takes approximately 1 minute on GPU, and estimating one pairwise brain-to-brain rotation takes approximately 5 minutes on CPU.
\begin{tabular}{llr|ccc}
\toprule
\textbf{Brain region} & \textbf{Preprocessing} & \textbf{Region size} 
& \textbf{Mean Rank} $\downarrow$ & \textbf{R@1} $\uparrow$ & \textbf{RSA} $\uparrow$ \\
\midrule
V1 \cite{Glasser2016} & Volumetric 1mm & 19,820 voxels & 23.42 & 0.54 & 0.37 \\
V1--V4 \cite{Glasser2016} & Volumetric 1mm & 51,197 voxels & 8.20 & 0.74 & 0.47 \\
Kastner2015 \cite{Kastner2015} & Volumetric 1mm & 56,743 voxels & 6.05 & 0.80 & 0.51 \\
NSDGeneral \cite{Allen2022} & Volumetric 1mm & 84,848 voxels & 5.28 & 0.82 & 0.53 \\
Streams \cite{Allen2022} & Volumetric 1mm & 152,370 voxels & 5.18 & 0.82 & 0.53 \\
% Whole cortex\cite{Glasser2016} & Volumetric 1mm & -- voxels & -- & -- & -- \\ % <- Not enough RAM
\midrule
NSDGeneral \cite{Allen2022} & Native surface & 68,464 vertices & 5.59 & 0.81 & 0.52 \\
\bottomrule
\end{tabular}
\begin{tabular}{lccccccc}
\toprule
\textbf{Configuration} 
& \textbf{Reliability} & \textbf{PCA} & \textbf{MCCA} & \textbf{Nonlinear}
& \textbf{Mean Rank} $\downarrow$ & \textbf{R@1} $\uparrow$ & \textbf{RSA} $\uparrow$ \\
\midrule
PCA only & -- & $\checkmark$ & -- & -- & 77.31 & 0.20 & 0.27 \\
Reliability + PCA & $\checkmark$ & $\checkmark$ & -- & -- & 116.52 & 0.09 & 0.13 \\
PCA + MCCA & -- & $\checkmark$ & $\checkmark$ & -- & 77.31 & 0.20 & 0.27 \\
Reliability + PCA + MCCA & $\checkmark$ & $\checkmark$ & $\checkmark$ & -- & 11.73 & 0.73 & 0.23 \\
\midrule
PCA + nonlinear & -- & $\checkmark$ & -- & $\checkmark$ & 14.05 & 0.60 & 0.55 \\
PCA + MCCA + nonlinear & -- & $\checkmark$ & $\checkmark$ & $\checkmark$ & 14.05 & 0.60 & 0.54 \\
Reliability + PCA + nonlinear & $\checkmark$ & $\checkmark$ & -- & $\checkmark$ & 21.89 & 0.43 & 0.51 \\
Full encoder & $\checkmark$ & $\checkmark$ & $\checkmark$ & $\checkmark$ & \textbf{5.28} & \textbf{0.82} & \textbf{0.53} \\
\bottomrule
\end{tabular}
\begin{tabular}{l|cccccccc|c}
\toprule
\textbf{Method}
& \textbf{S1} & \textbf{S2} & \textbf{S3} & \textbf{S4}
& \textbf{S5} & \textbf{S6} & \textbf{S7} & \textbf{S8}
& \textbf{Avg.} \\
\midrule
Preprocessed fMRI (GLMsingle~\cite{GLMSingle2022})
& 114.48 & 122.62 & 155.37 & 154.53 & 118.40 & 150.04 & 155.55 & 182.19 & 144.15 \\
\midrule
PCA
& 54.28 & 35.95 & 98.65 & 83.27 & 68.03 & 86.60 & 93.09 & 104.93 & 78.10 \\
PCA + SplitAE~\cite{SplitAE}
& 60.99 & 58.23 & 104.29 & 101.37 & 65.04 & 105.20 & 113.08 & 142.11 & 93.78 \\
PCA + DeepCCA~\cite{DeepCCA}
& 1.94 & 4.72 & 31.67 & 22.27 & 15.75 & 22.20 & 30.93 & 56.18 & 23.21 \\
PCA + Multiview CCA~\cite{RegMCCA2011}
& 1.40 & 2.79 & 27.95 & 19.78 & 9.57 & 17.71 & 19.72 & 54.92 & 19.23 \\
\midrule
DINOv2 ViT-S/14~\cite{dinov2} + Ridge
& 11.43 & 16.42 & 28.99 & 26.76 & 13.12 & 20.75 & 29.6 & 51.52 & 24.82 \\
CLIP ViT-B/16~\cite{Radford2021} + Ridge
& 12.3 & 16.88 & 27.1 & 24.86 & 11.71 & 18.1 & 28.25 & 46.66 & 23.23 \\
ViT AugReg-L/16~\cite{steiner2022how} + Ridge
& 7.93 & 12.9 & 24.51 & 22.98 & 10.22 & 15.4 & 24.37 & 44.12 & 20.30 \\
all-MiniLM-L6-v2~\cite{SentenceBert} + Ridge
& 18.67 & 24.8 & 35.63 & 31.97 & 18.24 & 25.45 & 34.0 & 55.00 & 30.47 \\
\midrule
Ours (linear)
& 1.09	& 1.20 & 20.55 & 13.4 & 4.54 & 8.42 & 15.38 & 29.26 & 11.73 \\
Ours (full)
& \textbf{1.03} & \textbf{1.12} & \textbf{11.25} & \textbf{4.69} & \textbf{1.77} & \textbf{3.09} & \textbf{7.60} & \textbf{11.67} & \textbf{5.28} \\
\bottomrule
\end{tabular}
\begin{tabular}{l|cccccccc|c}
\toprule
\textbf{Model}
& \textbf{S1} & \textbf{S2} & \textbf{S3} & \textbf{S4}
& \textbf{S5} & \textbf{S6} & \textbf{S7} & \textbf{S8}
& \textbf{Avg.} \\
\midrule
\multicolumn{10}{l}{\textit{Vision models}} \\
\midrule
DINOv2 ViT-S/14 \cite{dinov2} & 11.43 & 16.42 & 28.99 & 26.76 & 13.12 & 20.75 & 29.60 & 51.52 & 24.82 \\
DINOv2 ViT-B/14  & 19.11 & 24.55 & 35.68 & 38.05 & 20.33 & 32.07 & 40.18 & 65.86 & 34.48 \\
DINOv2 ViT-L/14 & 18.57 & 24.35 & 38.20 & 39.97 & 19.70 & 32.75 & 42.43 & 73.26 & 36.15 \\
DINOv2 ViT-G/14 & 26.51 & 30.49 & 44.01 & 46.19 & 24.07 & 39.55 & 49.98 & 77.67 & 42.31 \\
\midrule
AugReg ViT-T/16 \cite{steiner2022how} & 14.33 & 17.77 & 29.29 & 27.42 & 14.32 & 20.53 & 30.00 & 48.51 & 25.27 \\
AugReg ViT-S/16 & 11.87 & 17.66 & 27.84 & 26.57 & 12.26 & 19.31 & 29.49 & 48.57 & 24.20 \\
AugReg ViT-B/16 & 8.63 & 13.95 & 25.15 & 23.07 & 10.99 & 16.27 & 25.25 & 43.62 & 20.87 \\
AugReg ViT-L/16 & \textbf{7.93} & \textbf{12.90} & \textbf{24.51} & \textbf{22.98} & \textbf{10.22} & \textbf{15.40} & \textbf{24.37} & \textbf{44.12} & \textbf{20.30} \\
\midrule
MAE ViT-B/16 \cite{mae} & 18.67 & 25.38 & 40.77 & 35.04 & 20.81 & 30.20 & 41.84 & 59.57 & 34.03 \\
MAE ViT-L/16 & 12.63 & 17.72 & 32.39 & 27.66 & 14.20 & 22.66 & 33.31 & 50.08 & 26.33 \\
MAE ViT-H/14 & 13.23 & 18.17 & 32.98 & 27.89 & 14.12 & 23.25 & 34.80 & 49.73 & 26.77 \\
\midrule
CLIP ViT-B/16 FT \cite{Radford2021} & 17.97 & 23.00 & 34.80 & 34.59 & 17.65 & 26.87 & 36.85 & 61.08 & 31.60 \\
CLIP ViT-L/14 FT & 14.08 & 20.36 & 31.19 & 30.23 & 14.18 & 23.15 & 33.49 & 56.79 & 27.93 \\
CLIP ViT-H/14 FT & 16.70 & 23.27 & 35.24 & 33.82 & 16.89 & 26.59 & 38.52 & 62.65 & 31.71 \\
\midrule
CLIP ViT-B/16 \cite{Radford2021} & 12.30 & 16.88 & 27.10 & 24.86 & 11.71 & 18.10 & 28.25 & 46.66 & 23.23 \\
CLIP ViT-L/14 & 12.78 & 17.99 & 28.57 & 25.85 & 12.36 & 19.14 & 29.21 & 48.99 & 24.36 \\
CLIP ViT-H/14 & 10.86 & 15.74 & 26.51 & 23.78 & 11.36 & 17.54 & 26.88 & 46.87 & 22.44 \\
\midrule
\multicolumn{10}{l}{\textit{Language models}} \\
\midrule
E5-small & 24.19 & 30.91 & 42.11 & 38.52 & 21.75 & 31.60 & 42.15 & 65.67 & 37.11 \\
E5 & 22.92 & 30.18 & 41.78 & 38.04 & 20.71 & 31.02 & 41.41 & 65.26 & 36.41\\
Stella & 22.34 & 29.00 & 39.95 & 36.62 & 19.86 & 29.19 & 39.28 & 63.24 & 34.93 \\
Granite Embedding-small & 22.50 & 27.69 & 38.77 & 34.96 & 19.04 & 29.05 & 38.55 & 60.79 & 33.92 \\
Granite Embedding & 21.78 & 27.77 & 38.80 & 34.80 & 19.38 & 28.58 & 37.63 & 61.37 & 33.76\\
GTR & 19.49 & 24.90 & 36.21 & 33.78 & 16.95 & 26.84 & 35.83 & 58.75 & 31.59\\
all-MiniLM-L6-v2 & 18.67 & 24.80 & 35.63 & 31.97 & 18.24 & 25.45 & 34.00 & 55.00 & 30.47 \\
\bottomrule
\end{tabular}