Reproduce “How Does Muon Work in Private Deep Learning? Scalable Differentially Private Training via Orthogonalization” 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
Alexander Gaponov, Grigory Malinovsky, Egor Shulgin, Peter Richtarik
The authors study DP-Muon, applying Muon’s scale-invariant update to clipped, noise-added gradients. Since positive rescaling leaves the update unchanged, their analysis bounds a variance-dependent directional error and gives a K^-1/4 convergence rate toward a clipping-dependent neighborhood, but proves no improvement over clipped SGD. Results depend on batch size: at 1,024 and ε=8, DP-Adam reports 6.33 nats versus DP-Muon’s 6.52.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.