AI RESEARCH
Shuffling the Data, Stretching the Step-size: Sharper Bias in constant step-size SGD
arXiv CS.LG
•
ArXi:2604.10373v1 Announce Type: cross From adversarial robustness to multi-agent learning, many machine learning tasks can be cast as finite-sum min-max optimization or, generally, as variational inequality problems (VIPs). Owing to their simplicity and scalability, stochastic gradient methods with constant step size are widely used, despite the fact that they converge only up to a constant term.