Reproduce “Mask-Aware Execution for Efficient JEPA Training” 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

Md Musfiqur Rahman Sanim, Zhihao Shu, Bahram Afsharmanesh, Amirali Mirian, Wei Niu, Gagan Agrawal

M-JEPA accelerates multi-mask JEPA training by changing execution rather than the learning objective. It reuses mask-independent computations, routes mask-specific tokens together, and encodes only required target tokens. On A100 GPUs, authors report V-JEPA training-latency speedups from 1.1× at 2 masks to 1.7× at 10, although a separate ablation reports 1.45× at 10; memory use increases.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.