Michael Crawshaw

Non-Euclidian Gradient Descent Occurs at the Edge of Stability

An Exploration of Non-Euclidean Gradient Descent: Muon and its Many Variants

Provable Weak-to-Strong Generalization via Overspecified Students and Underspecified Teachers

Local Steps Speed Up Local GD for Heterogeneous Distributed Logistic Regression

Complexity Lower Bounds of Adaptive Gradient Algorithms for Non-convex Stochastic Optimization under Relaxed Smoothness

Constant Stepsize Local GD for Logistic Regression: Acceleration by Instability