MARINA: Faster non-convex distributed learning with compression
We develop and analyze MARINA: a new communication efficient method for non-convex distributed learning over heterogeneous datasets. MARINA employs a novel communication compression strategy based on the compression of gradient differences that is reminiscent of but different from the strategy emplo...
Saved in:
Main Authors: | , , , |
---|---|
Format: | text |
Language: | English |
Published: |
Institutional Knowledge at Singapore Management University
2021
|
Subjects: | |
Online Access: | https://ink.library.smu.edu.sg/sis_research/8682 https://ink.library.smu.edu.sg/context/sis_research/article/9685/viewcontent/ICML21_full_marina.pdf |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Singapore Management University |
Language: | English |
id |
sg-smu-ink.sis_research-9685 |
---|---|
record_format |
dspace |
spelling |
sg-smu-ink.sis_research-96852024-03-28T09:04:50Z MARINA: Faster non-convex distributed learning with compression GORBUNOV, Eduard BURLACHENKO, Konstantin LI, Zhize RICHTARIK, Peter We develop and analyze MARINA: a new communication efficient method for non-convex distributed learning over heterogeneous datasets. MARINA employs a novel communication compression strategy based on the compression of gradient differences that is reminiscent of but different from the strategy employed in the DIANA method of Mishchenko et al. (2019). Unlike virtually all competing distributed first-order methods, including DIANA, ours is based on a carefully designed biased gradient estimator, which is the key to its superior theoretical and practical performance. The communication complexity bounds we prove for MARINA are evidently better than those of all previous first-order methods. Further, we develop and analyze two variants of MARINA: VR-MARINA and PP-MARINA. The first method is designed for the case when the local loss functions owned by clients are either of a finite sum or of an expectation form, and the second method allows for a partial participation of clients -- a feature important in federated learning. All our methods are superior to previous state-of-the-art methods in terms of oracle/communication complexity. Finally, we provide a convergence analysis of all methods for problems satisfying the Polyak-Lojasiewicz condition. 2021-07-01T07:00:00Z text application/pdf https://ink.library.smu.edu.sg/sis_research/8682 https://ink.library.smu.edu.sg/context/sis_research/article/9685/viewcontent/ICML21_full_marina.pdf http://creativecommons.org/licenses/by-nc-nd/4.0/ Research Collection School Of Computing and Information Systems eng Institutional Knowledge at Singapore Management University Databases and Information Systems |
institution |
Singapore Management University |
building |
SMU Libraries |
continent |
Asia |
country |
Singapore Singapore |
content_provider |
SMU Libraries |
collection |
InK@SMU |
language |
English |
topic |
Databases and Information Systems |
spellingShingle |
Databases and Information Systems GORBUNOV, Eduard BURLACHENKO, Konstantin LI, Zhize RICHTARIK, Peter MARINA: Faster non-convex distributed learning with compression |
description |
We develop and analyze MARINA: a new communication efficient method for non-convex distributed learning over heterogeneous datasets. MARINA employs a novel communication compression strategy based on the compression of gradient differences that is reminiscent of but different from the strategy employed in the DIANA method of Mishchenko et al. (2019). Unlike virtually all competing distributed first-order methods, including DIANA, ours is based on a carefully designed biased gradient estimator, which is the key to its superior theoretical and practical performance. The communication complexity bounds we prove for MARINA are evidently better than those of all previous first-order methods. Further, we develop and analyze two variants of MARINA: VR-MARINA and PP-MARINA. The first method is designed for the case when the local loss functions owned by clients are either of a finite sum or of an expectation form, and the second method allows for a partial participation of clients -- a feature important in federated learning. All our methods are superior to previous state-of-the-art methods in terms of oracle/communication complexity. Finally, we provide a convergence analysis of all methods for problems satisfying the Polyak-Lojasiewicz condition. |
format |
text |
author |
GORBUNOV, Eduard BURLACHENKO, Konstantin LI, Zhize RICHTARIK, Peter |
author_facet |
GORBUNOV, Eduard BURLACHENKO, Konstantin LI, Zhize RICHTARIK, Peter |
author_sort |
GORBUNOV, Eduard |
title |
MARINA: Faster non-convex distributed learning with compression |
title_short |
MARINA: Faster non-convex distributed learning with compression |
title_full |
MARINA: Faster non-convex distributed learning with compression |
title_fullStr |
MARINA: Faster non-convex distributed learning with compression |
title_full_unstemmed |
MARINA: Faster non-convex distributed learning with compression |
title_sort |
marina: faster non-convex distributed learning with compression |
publisher |
Institutional Knowledge at Singapore Management University |
publishDate |
2021 |
url |
https://ink.library.smu.edu.sg/sis_research/8682 https://ink.library.smu.edu.sg/context/sis_research/article/9685/viewcontent/ICML21_full_marina.pdf |
_version_ |
1795302170802782208 |