Reproduce “A Stochastic Latent Variable Generalization of Joint-Embedding Predictive Architectures” 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

Nikos Giakoumoglou, Paschalis Giakoumoglou, Tania Stathaki

S-JEPA changes masked-image prediction by sampling a latent code for plausible hidden-region embeddings rather than reconstructing pixels. During training, a posterior sees context and target, while an inference-time prior uses context alone; a KL penalty aligns their codes. On ImageNet-1K frozen-feature linear evaluation, ViT-L/16 S-JEPA scores 78.3%, versus I-JEPA’s 77.5% after 600 epochs, a single-run result.

Summary by alphaXiv.

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