ICML 2020poster59 citations

A new regret analysis for Adam-type algorithms

Ahmet Alacaoglu, Yura Malitsky, Panayotis Mertikopoulos, Volkan Cevher

Abstract

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 we 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.

BibTeX
@InProceedings{pmlr-v119-alacaoglu20b,
  title = 	 {A new regret analysis for {A}dam-type algorithms},
  author =       {Alacaoglu, Ahmet and Malitsky, Yura and Mertikopoulos, Panayotis and Cevher, Volkan},
  booktitle = 	 {Proceedings of the 37th International Conference on Machine Learning},
  pages = 	 {202--210},
  year = 	 {2020},
  editor = 	 {III, Hal Daumé and Singh, Aarti},
  volume = 	 {119},
  series = 	 {Proceedings of Machine Learning Research},
  month = 	 {13--18 Jul},
  publisher =    {PMLR},
  pdf = 	 {http://proceedings.mlr.press/v119/alacaoglu20b/alacaoglu20b.pdf},
  url = 	 {https://proceedings.mlr.press/v119/alacaoglu20b.html},
  abstract = 	 {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 we 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.}
}
A new regret analysis for Adam-type algorithms · ICML 2020