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

Researchers showed that fMRI representations from different people’s visual cortex can be aligned using only their geometry, without paired training examples.

They trained subject-specific embeddings from repeated visual responses and used unsupervised rotation-based methods to map them across individuals.

The alignment recovered instance-level correspondences with high accuracy, suggesting the neural spaces are nearly isometric across subjects.

This could improve cross-subject neuroscience analysis and brain decoding by reducing the need for labeled matches.