Social Twitter Jan 4
@_arohan_ I think you missed the main ideas. - The basic premise of JEPA is that training by reconst...
By @ylecun
95 score
AI Analysis
Yann LeCun provides detailed technical explanation of JEPA (Joint Embedding Predictive Architecture), explaining why prediction should happen in representation space, methods to prevent collapse (EMA vs Infomax), and why he believes dimension-contrastive methods like SIGReg/LeJEPA are the future of self-supervised learning.
@_arohan_ I think you missed the main ideas.
- The basic premise of JEPA is that training by reconstructio/prediction in input space is evil (or counterproductive). The details are almost always unpredictable. Hence prediction must take place in representation space, where unpredictable details are eliminated.
- The main issue with JEPA is how to prevent collapse (in the absence of reconstruction loss). There are two classes of methods:
self-supervised learningJEPArepresentation learningtechnical deep-dive
Source details
@_arohan_ I think you missed the main ideas.
- The basic premise of JEPA is that training by reconstructio/prediction in input space is evil (or counterproductive). The details are almost always unpredictable. Hence prediction must take place in representation space, where unpredictable details are eliminated.
- The main issue with JEPA is how to prevent collapse (in the absence of reconstruction loss). There are two classes of methods:
self-supervised learningJEPArepresentation learningtechnical deep-dive