A new regret analysis for Adam-type algorithms

In this paper, we focus on a theory-practice gap for Adam and its variants (AMSgrad, AdamNC, etc.). In practice, these algorithms are used with a constant first-order moment parameter $\beta_{1}$ (typically between $0.9$ and $0.99$). In theory, regret guarantees for online convex optimization require a rapidly decaying $\beta_{1}\to0$ schedule. We show that this is an artifact of the standard analysis and propose a novel framework that allows us to derive optimal, data-dependent regret bounds with a constant $\beta_{1}$, without further assumptions. We also demonstrate the flexibility of our analysis on a wide range of different algorithms and settings.

[1]  Diego Klabjan,et al.  Convergence Analyses of Online ADAM Algorithm in Convex Setting and Two-Layer ReLU Neural Network , 2019, ArXiv.

[2]  Elad Hazan,et al.  Introduction to Online Convex Optimization , 2016, Found. Trends Optim..

[3]  Bin Dong,et al.  Nostalgic Adam: Weighing more of the past gradients when designing the adaptive learning rate , 2019, IJCAI.

[4]  Francesco Orabona,et al.  On the Convergence of Stochastic Gradient Descent with Adaptive Stepsizes , 2018, AISTATS.

[5]  Jinghui Chen,et al.  Closing the Generalization Gap of Adaptive Gradient Methods in Training Deep Neural Networks , 2018, IJCAI.

[6]  David E. Cox,et al.  ZO-AdaMM: Zeroth-Order Adaptive Momentum Method for Black-Box Optimization , 2019, NeurIPS.

[7]  Lijun Zhang,et al.  SAdam: A Variant of Adam for Strongly Convex Functions , 2019, ICLR.

[8]  Jimmy Ba,et al.  Adam: A Method for Stochastic Optimization , 2014, ICLR.

[9]  Li Shen,et al.  A Sufficient Condition for Convergences of Adam and RMSProp , 2018, 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR).

[10]  Mingyi Hong,et al.  On the Convergence of A Class of Adam-Type Algorithms for Non-Convex Optimization , 2018, ICLR.

[11]  Yoram Singer,et al.  Adaptive Subgradient Methods for Online Learning and Stochastic Optimization , 2011, J. Mach. Learn. Res..

[12]  Claudio Gentile,et al.  Adaptive and Self-Confident On-Line Learning Algorithms , 2000, J. Comput. Syst. Sci..

[13]  Pedro Savarese.  On the Convergence of AdaBound and its Connection to SGD , 2019, ArXiv.

[14]  Xu Sun,et al.  Adaptive Gradient Methods with Dynamic Bound of Learning Rate , 2019, ICLR.

[15]  Yuan Cao,et al.  On the Convergence of Adaptive Gradient Methods for Nonconvex Optimization , 2018, ArXiv.

[16]  Elad Hazan,et al.  Logarithmic regret algorithms for online convex optimization , 2006, Machine Learning.

[17]  Sanjiv Kumar,et al.  On the Convergence of Adam and Beyond , 2018 .