Physic Latent JEPA

Research prototype built in 2 days at the World Model Hackathon hosted by Yann LeCun: a self-supervised JEPA trained on the TUAB clinical EEG corpus (2,717 recordings), evaluated with strictly patient-disjoint probes.

Core idea: physics anchors (amplitude, phase, bandpower) tie the latent to physiological meaning. Hypothesis: purely geometric regularizers (VICReg-style) prevent trivial collapse but still allow functional collapse — a latent that is healthy geometrically yet physically meaningless.

Reached 79.0% balanced accuracy on normal/abnormal classification, beating masked, temporal and random baselines, with the healthiest latent of all variants: effective rank 14.8, 0% dead dimensions.

A physics-informed VAE augmentation benchmark added +3.6 to +4.7 balanced-accuracy points at 5–100% label fractions, while classical augmentation hurt performance — supporting the latent's sensitivity to spectral structure.