Research map / Label-efficient and robust learning / Recognition and long-tail

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

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Contrastive representation theory · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Learns subject-specific fMRI embeddings from repeated NSD stimulus presentations with a self-supervised MCCA+MLP encoder, then translates subjects into a shared space via unsupervised orthogonal rotations and spectral synchronization.

Problem

Whether shared neural geometry can be recovered across human brains from independently learned fMRI embeddings alone, without paired cross-subject data or model-derived anchors.

Approach

Extends the Strong Platonic Representation Hypothesis to brains: repetition-invariant self-supervised subject encoders (reliability weighting, PCA, MCCA, residual MLP with InfoNCE) followed by mini-vec2vec-style unsupervised orthogonal Procrustes/ICP pairwise maps, synchronized spectrally into one shared space.

Key results

• Unsupervised orthogonal brain-to-brain translation reaches Mean Rank 2.56 and R@1 0.78 averaged over ordered subject pairs (chance rank 258), beating vision and language model-guided ridge baselines on retrieval.
• Spectral synchronization into a shared space improves average Mean Rank from 2.47 to 1.97 and R@1 from 0.79 to 0.83 over pairwise translations.
• Within-subject encoder achieves Mean Rank 5.28 and R@1 0.82 averaged over 8 subjects, versus 19.23 and 0.60 for multiview CCA.

Surprising findings

• Supervised semi-orthogonal and ridge mappings from model embeddings to brain embeddings underperform unsupervised brain-to-brain orthogonal translation, suggesting model spaces are not isometric to brain spaces.
• Averaging the 10 seed rotations without seed selection degrades Mean Rank (4.61 vs 2.56), so results depend on best-seed selection.

Limitations

• Relies on high-SNR fMRI with repeated presentations; unclear for lower-SNR or smaller datasets.
• Unpaired translation is sensitive to initialization; best-seed selection uses held-out evaluation and no internal selection criterion is provided.
• Evaluated only on NSD visual cortex with a single 515-image shared set and 8 subjects.

Experimental setup

Datasets: Natural Scenes Dataset COCO

Benchmarks: Natural Scenes Dataset shared-image retrieval (515 images)

Metrics: Mean Rank R@1 RSA

Baselines: PCA SplitAE DeepCCA Multiview CCA GLMsingle DINOv2 ViT-S/14 + Ridge CLIP ViT-B/16 + Ridge ViT AugReg-L/16 + Ridge all-MiniLM-L6-v2 + Ridge Entropic GW

Builds on: Platonic Representation Hypothesis mini-vec2vec multi-view CCA hyperalignment

Closest papers by meaning

• FlatClip: Reusing Image Foundation Models for fMRI Representation Learning via Cortical Flatmaps (NeurIPS 2026 · similarity 0.88)
• Displacement Geometry Captures Platonic Shared Reality Across Models and Modalities (NeurIPS 2026 · similarity 0.87)
• Position: The Platonic Representation Hypothesis (ICML 2024 · similarity 0.87)
• Bridging Critical Gaps in Convergent Learning: How Representational Alignment Evolves Across Layers, Training, and Distribution Shifts (NeurIPS 2025 · similarity 0.87)
• Latent Functional Maps: a spectral framework for representation alignment (NeurIPS 2024 · similarity 0.86)
• Geometry-Aware Similarity Metrics for Neural Representations on Riemannian and Statistical Manifolds (NeurIPS 2026 · similarity 0.86)