SplitJEPA paper separates invariant and variant factors in JEPA latent states
mayfer · x · 2026-10-11
- SplitJEPA is a new paper that combines predictive learning with matched observations sharing invariant factors but differing in variant factors, using their complementary information to resolve latent block ambiguity in JEPA representations.
- The motivating example: a mug looks different under morning light vs. dim night light, but your hand still reaches the same spot — the brain keeps spatial location separate from lighting, and SplitJEPA gives JEPA the same separation.
- The post notes the LeJEPA line of work (LeCun, Balestriero, Klindt) already showed that under stationary Gaussian predictive dynamics, latent prediction with Gaussian regularization recovers the complete latent state up to a global orthogonal transform, with no reconstruction needed; SplitJEPA targets removing that remaining ambiguity.
Related event: SplitJEPA Separates Invariant and Variable Factors Without Reconstruction(2 posts)→
More from Research
- LeaP from Southeast University, accepted at CoRL 2026, makes diffusion action priors state-aware — jiqizhixin · 2026-10-11
- DeformX co-simulation framework for deformable objects lands IROS 2026 oral — rsasaki0109 · 2026-10-11
- Independent Researcher Makes TPU Pallas top-k Bitwise Correct and 1.67x Faster — Francis_YAO_ · 2026-10-11
- Eric Jang resurfaces 2021 essay: bet on compute pressure sparking spontaneous intelligence — ericjang11 · 2026-10-11
- New arXiv paper examines scaling and emergent abstractions in byte-level language models — yogthos · 2026-10-11
- NVIDIA's GATOR turns casual photos into simulation-ready 3D objects with agentic refinement — AjayMandlekar · 2026-10-11