Radar ·
SIGReg and repulsion in JEPA training
A September 28, 2026 post describes how SIGReg, used by LeJEPA and LeWM, approximates a sliced MMD whose expansion reveals repulsion between pairs of samples in a batch.
A post dated September 28, 2026 says SIGReg, used by LeJEPA and LeWM, approximates a sliced MMD. According to the post, expanding this approximation reveals a repulsion effect between pairs of samples in a batch during JEPA training.
The finding offers a way to interpret how a regularizer can shape representations in this training. To check the claim's scope, consult the original post and verify its mathematical formulation and evidence; the available summary gives no experiments or quantitative results. If you use AI to study or apply this material, avoid entering personal data or confidential documents.
Original source: SIGReg Adiciona Repulsão entre Pares ao Treinamento de JEPA