Yann LeCun's JEPA Explained

3-Second Blueprint (English)

1. Prevents representation collapse in self-supervised image embeddings by shaping the latent cloud into an isotropic Gaussian curve.
2. Applies the Cramér-Wold theorem to verify high-dimensional Gaussian geometry across random 1D projections without bin explosion.
3. Winds projected points onto a circle to match characteristic functions smoothly via SIGReg, enabling stable gradient optimization.