Unsupervised learning reveals universal brain geometry across individuals

A new paper on arXiv titled "Platonic Representations in the Human Brain: Unsupervised Recovery of Universal Geometry" tests whether the Strong Platonic Representation Hypothesis—previously applied to artificial neural networks—also holds for human brains. The hypothesis suggests that neural representations converge to a universal latent space even when learned independently. To test this, the researchers used fMRI recordings from the Natural Scenes Dataset, which contains repeated stimulus presentations across subjects. They trained a self-supervised encoder to learn subject-specific embeddings purely from brain activity, without any paired cross-subject data or intermediate model representations.

The team then translated these independently learned embeddings across subjects using unsupervised orthogonal rotations. Remarkably, they successfully aligned the subject-specific spaces into a common coordinate system, and synchronizing pairwise rotations further improved cross-subject retrieval accuracy. These results provide direct evidence for a shared neural geometry in the human visual cortex: subject-specific fMRI representations are approximately isometric across individuals and can be mapped via purely geometric transformations. This work bridges neuroscience and AI, suggesting that universal geometries exist not only in artificial networks but also in biological brains, opening new possibilities for cross-subject brain decoding and brain-computer interfaces.