Mini-batch primal and dual methods for SVMs

Martin Takáč; Avleen Bijral; Peter Richtárik; Nathan Srebro

Mini-batch primal and dual methods for SVMs

Martin Takáč, Avleen Bijral, Peter Richtárik, Nathan Srebro

Research output: Contribution to conference › Paper › peer-review

Abstract

We address the issue of using mini-batches in stochastic optimization of SVMs. We show that the same quantity, the spectral norm of the data, controls the parallelization speedup obtained for both primal stochastic subgradient descent (SGD) and stochastic dual coordinate ascent (SCDA) methods and use it to derive novel variants of mini-batched SDCA. Our guarantees for both methods are expressed in terms of the original nonsmooth primal problem based on the hinge-loss.

Original language	English (US)
Pages	2059-2067
Number of pages	9
State	Published - 2013
Externally published	Yes
Event	30th International Conference on Machine Learning, ICML 2013 - Atlanta, GA, United States Duration: Jun 16 2013 → Jun 21 2013

Other

Other	30th International Conference on Machine Learning, ICML 2013
Country/Territory	United States
City	Atlanta, GA
Period	06/16/13 → 06/21/13

ASJC Scopus subject areas

Human-Computer Interaction
Sociology and Political Science

Cite this

@conference{e96502fbf42447b58a2527dd37096290,

title = "Mini-batch primal and dual methods for SVMs",

abstract = "We address the issue of using mini-batches in stochastic optimization of SVMs. We show that the same quantity, the spectral norm of the data, controls the parallelization speedup obtained for both primal stochastic subgradient descent (SGD) and stochastic dual coordinate ascent (SCDA) methods and use it to derive novel variants of mini-batched SDCA. Our guarantees for both methods are expressed in terms of the original nonsmooth primal problem based on the hinge-loss.",

author = "Martin Tak{\'a}{\v c} and Avleen Bijral and Peter Richt{\'a}rik and Nathan Srebro",

year = "2013",

language = "English (US)",

pages = "2059--2067",

note = "30th International Conference on Machine Learning, ICML 2013 ; Conference date: 16-06-2013 Through 21-06-2013",

}

TY - CONF

T1 - Mini-batch primal and dual methods for SVMs

AU - Takáč, Martin

AU - Bijral, Avleen

AU - Richtárik, Peter

AU - Srebro, Nathan

PY - 2013

Y1 - 2013

N2 - We address the issue of using mini-batches in stochastic optimization of SVMs. We show that the same quantity, the spectral norm of the data, controls the parallelization speedup obtained for both primal stochastic subgradient descent (SGD) and stochastic dual coordinate ascent (SCDA) methods and use it to derive novel variants of mini-batched SDCA. Our guarantees for both methods are expressed in terms of the original nonsmooth primal problem based on the hinge-loss.

AB - We address the issue of using mini-batches in stochastic optimization of SVMs. We show that the same quantity, the spectral norm of the data, controls the parallelization speedup obtained for both primal stochastic subgradient descent (SGD) and stochastic dual coordinate ascent (SCDA) methods and use it to derive novel variants of mini-batched SDCA. Our guarantees for both methods are expressed in terms of the original nonsmooth primal problem based on the hinge-loss.

UR - http://www.scopus.com/inward/record.url?scp=84897543082&partnerID=8YFLogxK

M3 - Paper

AN - SCOPUS:84897543082

SP - 2059

EP - 2067

T2 - 30th International Conference on Machine Learning, ICML 2013

Y2 - 16 June 2013 through 21 June 2013

ER -

Mini-batch primal and dual methods for SVMs

Abstract

Other

ASJC Scopus subject areas

Other files and links

Fingerprint

Cite this