Methods for Decentralized Signal Processing With Big Data

Hoi To Wai, Anna Scaglione, Eric Moulines

Research output: Chapter in Book/Report/Conference proceedingChapter

Abstract

Decentralized signal processing methods offer an attractive solution to perform big data analytics by harnessing distributed databases, servers, and network resources in a flexible and adaptive fashion. Often, machine learning algorithms on big data are instances of high-dimensional constrained optimization problems, where the projection step comes at a significant computational cost. In this context, the goal of this chapter is to review recent advances in the solution and analysis of decentralized projection-free algorithms, focusing on both convex and nonconvex problems and comparing them with decentralized projective gradient methods. The popular machine learning problem of robust low rank matrix completion is considered to validate our theoretical claims numerically.

Original languageEnglish (US)
Title of host publicationCooperative and Graph Signal Processing
Subtitle of host publicationPrinciples and Applications
PublisherElsevier
Pages399-417
Number of pages19
ISBN (Electronic)9780128136782
ISBN (Print)9780128136775
DOIs
StatePublished - Jun 20 2018
Externally publishedYes

Keywords

  • Big data analytics
  • Decentralized signal processing
  • Frank-Wolfe algorithm
  • Matrix completion
  • Projection-free optimization

ASJC Scopus subject areas

  • Medicine (miscellaneous)

Fingerprint Dive into the research topics of 'Methods for Decentralized Signal Processing With Big Data'. Together they form a unique fingerprint.

Cite this