Reproduce “Shaping Useful Noise: Energy Distributions Predict Visual Pretraining Quality” 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

Ching Lam Choi, Antonio Torralba, Phillip Isola, Stefanie Jegelka

The authors introduce Energy-Level Distribution Distance (ELD), which compares how real and synthetic images occupy likelihood levels under the same normalized real-image density model. Across nine synthetic sources in DINO/ViT-S pretraining, lower ELD is associated with higher frozen ImageNet-1K linear-probe accuracy (Spearman ρ = −0.87). This observational association does not establish causality and is limited to the studied settings.

Summary by alphaXiv.

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