学术文献库

Recently, Stochastic Gradient Descent (SGD) and its variants have become the dominant methods in the large-scale optimization of machine learning (ML) problems. A variety of strategies have been proposed for tuning the step sizes, ranging from adaptive step sizes to heuristic methods to change the step size in each iteration. Also, momentum has been widely employed in ML tasks to accelerate the training process. Yet, there is a gap in our theoretical understanding of them. In this work, we start to close this gap by providing formal guarantees to a few heuristic optimization methods and proposing improved algorithms. First, we analyze a generalized version of the AdaGrad (Delayed AdaGrad) step sizes in both convex and non-convex settings, showing that these step sizes allow the algorithms to automatically adapt to the level of noise of the stochastic gradients. We show for the first time sufficient conditions for Delayed AdaGrad to achieve almost sure convergence of the gradients to zero. Moreover, we present a high probability analysis for Delayed AdaGrad and its momentum variant in the non-convex setting. Second, we analyze SGD with exponential and cosine step sizes, which are empirically successful but lack theoretical support. We provide the very first convergence guarantees for them in the smooth and non-convex setting, with and without the Polyak-Łojasiewicz (PL) condition. We also show their good property of adaptivity to noise under the PL condition. Third, we study the last iterate of momentum methods. We prove the first lower bound in the convex setting for the last iterate of SGD with constant momentum. Moreover, we investigate a class of Follow-The-Regularized-Leader-based momentum algorithms with increasing momentum and shrinking updates. We show that their last iterate has optimal convergence for unconstrained convex stochastic optimization problems.