1/7 New paper accepted as ICML spotlight arxiv.org/abs/2605.03517! We unify self-supervised learning (SSL) algorithms (e.g., contrastive, VICReg, stopgrad) via latent distribution matching (LDM), which matches an induced latent distribution to an explicit latent model.
7/ Based on these guarantees, we argue that SSL’s success is not primarily based on discarding “irrelevant” information, but instead on constraining representations directly in the latent space.