Reproduce “Last-Iterate Convergence of Stochastic Gradient Methods for Unbounded Problems” 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
Kevin Kurian Thomas Vaidyan, Ahmet Alacaoglu, Michael P. Friedlander
The paper analyzes proximal stochastic gradient descent (SGD) to guarantee the final iterate rather than an average. It controls stochastic-gradient growth along the algorithm’s trajectory under B-G or component-wise convexity and smoothness (CW), then uses reference points to bound the final objective gap. Under B-G, a horizon-dependent schedule gives an expected gap of O(1/√(T+1)); the standard square-root schedule is suboptimal.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.