Reproduce “Asymmetric Multimodal Learning with Decomposed JEPAs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yusuf Shihata
D-JEPA separates image and text representations into an aligned shared code and private fibers, which preserve information through contextual prediction and can coordinate. The authors show that exact alignment bounds the representation’s information about the world by modality-redundant information. On CLEVR, D-JEPA reaches 39.61% joint-probe accuracy versus 21.25% for alignment-only, but remains below the frozen-backbone score of 45.57%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.