Restart and Adaptive Acceleration in Stochastic Gradient Methods
We study restart schemes in stochastic optimization problems for non-smooth and weakly convex that satisfy a Kurdyka-\L ojasiewicz inequality. We show that using restarts allows us to leverage the K{\L} inequalities to achieve improved rates of convergence, with acceleration depending explicitly on the K{\L} exponent. Furthermore, optimal restart schedules lead to learning-rates akin to Polyak steps for SGD. While regularity constants such as the K{\L} exponent are typically unknown in practice, we prove that restart schemes are robust to a significant misspecification of these constants, hence nearly adaptive. We detail numerical experiments on both toy problems, where the K{\L} exponent is controlled, and training of Large Language Models (LLMs).