On the Last Iterate Convergence of Momentum Methods

Xiaoyu Li, Mingrui Liu, Francesco Orabona
Proceedings of The 33rd International Conference on Algorithmic Learning Theory, PMLR 167:699-717, 2022.

Abstract

SGD with Momentum (SGDM) is a widely used family of algorithms for large scale optimization of machine learning problems. Yet, when optimizing generic convex functions, no advantage is known for any SGDM algorithm over plain SGD. Moreover, even the most recent results require changes to the SGDM algorithms, like averaging of the iterates and a projection onto a bounded domain, which are rarely used in practice. In this paper, we focus on the convergence rate of the last iterate of SGDM. For the first time, we prove that for any constant momentum factor, there exists a Lipschitz and convex function for which the last iterate of SGDM suffers from a suboptimal convergence rate of $\Omega(\frac{\log T}{\sqrt{T}})$ after $T$ iterations. Based on this fact, we study a class of (both adaptive and non-adaptive) Follow-The-Regularized-Leader-based SGDM algorithms with \emph{increasing momentum} and \emph{shrinking updates}. For these algorithms, we show that the last iterate has optimal convergence $O(\frac{1}{\sqrt{T}})$ for unconstrained convex stochastic optimization problems without projections onto bounded domains nor knowledge of $T$. Further, we show a variety of results for FTRL-based SGDM when used with adaptive stepsizes. Empirical results are shown as well.

Cite this Paper

BibTeX
@InProceedings{pmlr-v167-li22a, title = {On the Last Iterate Convergence of Momentum Methods}, author = {Li, Xiaoyu and Liu, Mingrui and Orabona, Francesco}, booktitle = {Proceedings of The 33rd International Conference on Algorithmic Learning Theory}, pages = {699--717}, year = {2022}, editor = {Dasgupta, Sanjoy and Haghtalab, Nika}, volume = {167}, series = {Proceedings of Machine Learning Research}, month = {29 Mar--01 Apr}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v167/li22a/li22a.pdf}, url = {https://proceedings.mlr.press/v167/li22a.html}, abstract = {SGD with Momentum (SGDM) is a widely used family of algorithms for large scale optimization of machine learning problems. Yet, when optimizing generic convex functions, no advantage is known for any SGDM algorithm over plain SGD. Moreover, even the most recent results require changes to the SGDM algorithms, like averaging of the iterates and a projection onto a bounded domain, which are rarely used in practice. In this paper, we focus on the convergence rate of the last iterate of SGDM. For the first time, we prove that for any constant momentum factor, there exists a Lipschitz and convex function for which the last iterate of SGDM suffers from a suboptimal convergence rate of $\Omega(\frac{\log T}{\sqrt{T}})$ after $T$ iterations. Based on this fact, we study a class of (both adaptive and non-adaptive) Follow-The-Regularized-Leader-based SGDM algorithms with \emph{increasing momentum} and \emph{shrinking updates}. For these algorithms, we show that the last iterate has optimal convergence $O(\frac{1}{\sqrt{T}})$ for unconstrained convex stochastic optimization problems without projections onto bounded domains nor knowledge of $T$. Further, we show a variety of results for FTRL-based SGDM when used with adaptive stepsizes. Empirical results are shown as well.} }
Endnote
%0 Conference Paper %T On the Last Iterate Convergence of Momentum Methods %A Xiaoyu Li %A Mingrui Liu %A Francesco Orabona %B Proceedings of The 33rd International Conference on Algorithmic Learning Theory %C Proceedings of Machine Learning Research %D 2022 %E Sanjoy Dasgupta %E Nika Haghtalab %F pmlr-v167-li22a %I PMLR %P 699--717 %U https://proceedings.mlr.press/v167/li22a.html %V 167 %X SGD with Momentum (SGDM) is a widely used family of algorithms for large scale optimization of machine learning problems. Yet, when optimizing generic convex functions, no advantage is known for any SGDM algorithm over plain SGD. Moreover, even the most recent results require changes to the SGDM algorithms, like averaging of the iterates and a projection onto a bounded domain, which are rarely used in practice. In this paper, we focus on the convergence rate of the last iterate of SGDM. For the first time, we prove that for any constant momentum factor, there exists a Lipschitz and convex function for which the last iterate of SGDM suffers from a suboptimal convergence rate of $\Omega(\frac{\log T}{\sqrt{T}})$ after $T$ iterations. Based on this fact, we study a class of (both adaptive and non-adaptive) Follow-The-Regularized-Leader-based SGDM algorithms with \emph{increasing momentum} and \emph{shrinking updates}. For these algorithms, we show that the last iterate has optimal convergence $O(\frac{1}{\sqrt{T}})$ for unconstrained convex stochastic optimization problems without projections onto bounded domains nor knowledge of $T$. Further, we show a variety of results for FTRL-based SGDM when used with adaptive stepsizes. Empirical results are shown as well.
APA
Li, X., Liu, M. & Orabona, F.. (2022). On the Last Iterate Convergence of Momentum Methods. Proceedings of The 33rd International Conference on Algorithmic Learning Theory, in Proceedings of Machine Learning Research 167:699-717 Available from https://proceedings.mlr.press/v167/li22a.html.

Related Material