arXiv 2511.08544
LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics
By Randall Balestriero and Yann LeCun
Published 2025-11-11
Mindmap
Browse the paper's core ideas, clusters, and relationships in a structured outline.
Learning manipulable representations of the world and its dynamics is central to AI. Joint-Embedding Predictive Architectures (JEPAs) offer a promising blueprint, but lack of practical guidance and theory has led to ad-hoc R&D. We present a comprehensive theory of JEPAs and instantiate it in LeJEPA , a lean, scalable, and theoretically grounded training objective. First, we identify the isotropic Gaussian as the opt…