dc.contributor.author Mishchenko, Konstantin dc.contributor.author Gorbunov, Eduard dc.contributor.author Takáč, Martin dc.contributor.author Richtarik, Peter dc.date.accessioned 2019-05-28T14:07:40Z dc.date.available 2019-05-28T14:07:40Z dc.date.issued 2019-01-26 dc.identifier.uri http://hdl.handle.net/10754/653106 dc.description.abstract Training very large machine learning models requires a distributed computingapproach, with communication of the model updates often being the bottleneck.For this reason, several methods based on the compression (e.g., sparsificationand/or quantization) of the updates were recently proposed, including QSGD(Alistarh et al., 2017), TernGrad (Wen et al., 2017), SignSGD (Bernstein etal., 2018), and DQGD (Khirirat et al., 2018). However, none of these methodsare able to learn the gradients, which means that they necessarily suffer fromseveral issues, such as the inability to converge to the true optimum in thebatch mode, inability to work with a nonsmooth regularizer, and slowconvergence rates. In this work we propose a new distributed learningmethod---DIANA---which resolves these issues via compression of gradientdifferences. We perform a theoretical analysis in the strongly convex andnonconvex settings and show that our rates are vastly superior to existingrates. Our analysis of block-quantization and differences between $\ell_2$ and$\ell_\infty$ quantization closes the gaps in theory and practice. Finally, byapplying our analysis technique to TernGrad, we establish the first convergencerate for this method. dc.description.sponsorship The work of Peter Richtarik was supported by the KAUST baseline funding scheme. The work of Martin Takac was partially supported by the U.S. National Science Foundation, under award numbers NSF:CCF:1618717, NSF:CMMI:1663256 and NSF:CCF:1740796. dc.publisher arXiv dc.relation.url https://arxiv.org/pdf/1901.09269 dc.rights Archived with thanks to arXiv dc.title Distributed Learning with Compressed Gradient Differences dc.type Preprint dc.contributor.department Computer Science dc.contributor.department Computer Science Program dc.contributor.department Computer, Electrical and Mathematical Sciences and Engineering (CEMSE) Division dc.eprint.version Pre-print dc.contributor.institution Moscow Institute of Physics and Technology, Russian Federation dc.contributor.institution Lehigh University, USA dc.contributor.institution University of Edinburgh, United Kingdom dc.identifier.arxivid 1901.09269 kaust.person Mishchenko, Konstantin kaust.person Richtarik, Peter refterms.dateFOA 2019-05-28T14:08:12Z
﻿

Name:
1901.09269.pdf
Size:
1.905Mb
Format:
PDF
Description:
Preprint