Announcement
Researchers say JEPA can learn the map—but needs help keeping the details
The team reports Slow SVReg preserves task-relevant details while improving planning across five visual control tasks.
TLDR
Learning the layout is only half the job. Mengye Ren and colleagues report that SVReg produces smooth internal maps in JEPA, but can lose object information such as cube height. Their Slow SVReg adds channels trained to change slowly and a predictor that learns changes to retain those details. The team reports the highest planning success among tested methods in the hard settings of five visual control tasks.
Combined views
8.8K
9 Sources, first seen ago

Comments