MotionJEPA: Oxford-led team fixes JEPA world models' bias toward slow, simple features
randall_balestr · x · 2026-09-23
Researchers from Oxford, vivo, and others released MotionJEPA, addressing JEPA world models' inductive bias toward slow features that suppresses dynamic information. In Pong experiments, baselines reconstruct scoreboards and bats but miss the fast-moving ball, while MotionJEPA learns balanced static/dynamic representations. The DISReg regularizer predicts temporal difference image embeddings without pixel reconstruction, combining a static term and an unconstrained dynamic term to prevent collapse. Paper and code are open.
More from Research
- Story Imprinting Paper Finds AI Assistants Absorb Traits From Resembling Human Characters — OwainEvans_UK · 2026-09-23
- Yann LeCun: desk-reject papers that fail LLM watermarking detection — RexDouglass · 2026-09-23
- Pedro Domingos unveils Tensor Logic, an AI language unifying deep learning and symbolic AI — pmddomingos · 2026-09-23
- Stanford's Bayesian pulse deconvolution extracts accurate heart signals from wearables — rbhar90 · 2026-09-23
- Researchers lament: no good RL environment sets exist that allow diffing behaviors across training runs — 1a3orn · 2026-09-23
- New PNAS paper: human learning is an understudied lever for boosting human–AI synergy — iyadrahwan · 2026-09-23