Stochastic Optimization with Heavy-Tailed Noise via Accelerated Gradient Clipping
Eduard Gorbunov, Marina Danilova, Alexander Gasnikov
Abstract
In this paper, we propose a new accelerated stochastic first-order method called clipped-SSTM for smooth convex stochastic optimization with heavy-tailed distributed noise in stochastic gradients and derive the first high-probability complexity bounds for this method closing the gap in the theory of stochastic optimization with heavy-tailed noise. Our method is based on a special variant of accelerated Stochastic Gradient Descent (SGD) and clipping of stochastic gradients. We extend our method to the strongly convex case and prove new complexity bounds that outperform state-of-the-art results in this case. Finally, we extend our proof technique and derive the first non-trivial high-probability complexity bounds for SGD with clipping without light-tails assumption on the noise.
BibTeX
@inproceedings{NEURIPS2020_abd1c782,
author = {Gorbunov, Eduard and Danilova, Marina and Gasnikov, Alexander},
booktitle = {Advances in Neural Information Processing Systems},
editor = {H. Larochelle and M. Ranzato and R. Hadsell and M.F. Balcan and H. Lin},
pages = {15042--15053},
publisher = {Curran Associates, Inc.},
title = {Stochastic Optimization with Heavy-Tailed Noise via Accelerated Gradient Clipping},
url = {https://proceedings.neurips.cc/paper_files/paper/2020/file/abd1c782880cc59759f4112fda0b8f98-Paper.pdf},
volume = {33},
year = {2020}
}