Non quia difficilia sunt non audemus, sed quia non audemus difficilia sunt
Home -> Publications
Home
  Publications
    
all years
    2019
    2018
    2017
    2016
    2015
    2014
    2013
    2012
    2011
    2010
    2009
    2008
    2007
    2006
    2005
    2004
    theses
    techreports
    presentations
    edited volumes
    conferences
  Awards
  Research
  Teaching
  BLOG
  Miscellaneous
  Full CV [pdf]






  Events








  Past Events





Publications of Torsten Hoefler
Copyright Notice:

The documents distributed by this server have been provided by the contributing authors as a means to ensure timely dissemination of scholarly and technical work on a noncommercial basis. Copyright and all rights therein are maintained by the authors or by other copyright holders, notwithstanding that they have offered their works here electronically. It is understood that all persons copying this information will adhere to the terms and constraints invoked by each author's copyright. These works may not be reposted without the explicit permission of the copyright holder.

Dan Alistarh, Torsten Hoefler, Mikael Johansson, Sarit Khirirat, Nikola Konstantinov, Cedric Renggli:

 The Convergence of Sparsified Gradient Methods

(In Advances in Neural Information Processing Systems 31, presented in Montreal, Canada, Curran Associates, Inc., Dec. 2018)

Abstract

Distributed training of massive machine learning models, in particular deep neural networks, via Stochastic Gradient Descent (SGD) is becoming commonplace. Several families of communication-reduction methods, such as quantization, largebatch methods, and gradient sparsification, have been proposed. To date, gradient sparsification methods–where each node sorts gradients by magnitude, and only communicates a subset of the components, accumulating the rest locally–are known to yield some of the largest practical gains. Such methods can reduce the amount of communication per step by up to three orders of magnitude, while preserving model accuracy. Yet, this family of methods currently has no theoretical justification. This is the question we address in this paper. We prove that, under analytic assumptions, sparsifying gradients by magnitude with local error correction provides convergence guarantees, for both convex and non-convex smooth objectives, for data-parallel SGD. The main insight is that sparsification methods implicitly maintain bounds on the maximum impact of stale updates, thanks to selection by magnitude. Our analysis and empirical validation also reveal that these methods do require analytical conditions to converge well, justifying existing heuristics.

Documents

download article:
 

BibTeX

@incollection{top-k,
  author={Dan Alistarh and Torsten Hoefler and Mikael Johansson and Sarit Khirirat and Nikola Konstantinov and Cedric Renggli},
  title={{The Convergence of Sparsified Gradient Methods}},
  year={2018},
  month={Dec.},
  booktitle={Advances in Neural Information Processing Systems 31},
  location={Montreal, Canada},
  publisher={Curran Associates, Inc.},
  source={http://www.unixer.de/~htor/publications/},
}

serving: 54.227.31.145:58432© Torsten Hoefler