Reproduce “When and How SignSGD Outperforms SGD: A Functional Scaling Law Perspective” 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
Zilin Wang, Binghui Li, Jia-Nan Wang, Lean Wang, Jinbo Wang, Lei Wu
The authors analyze SignSGD in a power-law kernel regression model, deriving a functional scaling law under label-noise-dominated dynamics. The theory predicts faster signal learning but slower noise forgetting than SGD, with task-dependent efficiency effects. In an easy synthetic task at data budget 32,000, measured critical batch sizes were 2,000 for SignSGD and 16 for SGD; these finite-grid results do not establish theoretical scaling exponents.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.