Ändra sökning
RefereraExporteraLänk till posten
Permanent länk

Direktlänk
Referera
Referensformat
  • apa
  • harvard1
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Annat format
Fler format
Språk
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Annat språk
Fler språk
Utmatningsformat
  • html
  • text
  • asciidoc
  • rtf
The Convergence of Sparsified Gradient Methods
IST Austria, Klosterneuburg, Austria..
Swiss Fed Inst Technol, Zurich, Switzerland..
KTH, Skolan för elektroteknik och datavetenskap (EECS), Reglerteknik.
KTH, Skolan för elektroteknik och datavetenskap (EECS), Reglerteknik.ORCID-id: 0000-0003-4473-2011
Visa övriga samt affilieringar
2018 (Engelska)Ingår i: Advances in Neural Information Processing Systems 31 (NIPS 2018) / [ed] Bengio, S Wallach, H Larochelle, H Grauman, K CesaBianchi, N Garnett, R, Neural Information Processing Systems (NIPS) , 2018, Vol. 31Konferensbidrag, Publicerat paper (Refereegranskat)
Abstract [en]

Stochastic Gradient Descent (SGD) has become the standard tool for distributed training of massive machine learning models, in particular deep neural networks. Several families of communication-reduction methods, such as quantization, large-batch methods, and gradient sparsification, have been proposed to reduce the overheads of distribution. 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 also reveals that these methods do require analytical conditions to converge well, justifying and complementing existing heuristics.

Ort, förlag, år, upplaga, sidor
Neural Information Processing Systems (NIPS) , 2018. Vol. 31
Serie
Advances in Neural Information Processing Systems, ISSN 1049-5258 ; 31
Nationell ämneskategori
Annan data- och informationsvetenskap
Identifikatorer
URN: urn:nbn:se:kth:diva-249918ISI: 000461852000047Scopus ID: 2-s2.0-85064812369OAI: oai:DiVA.org:kth-249918DiVA, id: diva2:1307203
Konferens
32nd Conference on Neural Information Processing Systems (NIPS), DEC 02-08, 2018, Montreal, Canada
Anmärkning

QC 20190426

Tillgänglig från: 2019-04-26 Skapad: 2019-04-26 Senast uppdaterad: 2019-10-09Bibliografiskt granskad

Open Access i DiVA

Fulltext saknas i DiVA

Scopus

Personposter BETA

Johansson, MikaelKhirirat, Sarit

Sök vidare i DiVA

Av författaren/redaktören
Johansson, MikaelKhirirat, Sarit
Av organisationen
Reglerteknik
Annan data- och informationsvetenskap

Sök vidare utanför DiVA

GoogleGoogle Scholar

urn-nbn

Altmetricpoäng

urn-nbn
Totalt: 205 träffar
RefereraExporteraLänk till posten
Permanent länk

Direktlänk
Referera
Referensformat
  • apa
  • harvard1
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Annat format
Fler format
Språk
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Annat språk
Fler språk
Utmatningsformat
  • html
  • text
  • asciidoc
  • rtf