Platonic Representations in the Human Brain: Unsupervised Recovery of Universal Geometry
Researchers have demonstrated that human brain representations of visual stimuli exhibit a shared underlying geometry. Using fMRI data and a self-supervised encoder, they learned subject-specific embeddings and showed these could be translated across individuals through unsupervised geometric transformations. This suggests that neural representations in the visual cortex are approximately isometric and compatible with a common coordinate system, potentially extending the concept of Platonic representations beyond artificial neural networks. AI
IMPACT Suggests potential for cross-subject brain data analysis and translation, extending AI representation concepts to neuroscience.