Skip to main navigation Skip to search Skip to main content

EE-Grad: Exploration and Exploitation for Cost-Efficient Mini-Batch SGD

  • University of Illinois at Urbana-Champaign
  • University of California at Irvine

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

We present a generic framework for mitigating the tradeoff between fidelity and cost in computing stochastic gradients when the costs of acquiring stochastic gradients of different quality are not known a priori. We consider a mini-batch oracle that distributes a limited query budget over a number of stochastic gradients and aggregates them to estimate the true gradient. Since the optimal mini-batch size depends on the unknown cost-fidelity function, we propose an algorithm, EE-Grad, that sequentially explores the performance of mini-batch oracles and exploits the accumulated knowledge to estimate the one achieving the best performance in terms of cost-efficiency. We provide performance guarantees for EE-Grad with respect to the optimal mini-batch oracle, and illustrate these results in the case of strongly convex objectives. We also provide a simple numerical example that corroborates our theoretical findings.

Original languageEnglish
Title of host publication55th Asilomar Conference on Signals, Systems and Computers, ACSSC 2021
EditorsMichael B. Matthews
PublisherIEEE Computer Society
Pages490-497
Number of pages8
ISBN (Electronic)9781665458283
DOIs
StatePublished - 2021
Event55th Asilomar Conference on Signals, Systems and Computers, ACSSC 2021 - Virtual, Pacific Grove, United States
Duration: Oct 31 2021Nov 3 2021

Publication series

NameConference Record - Asilomar Conference on Signals, Systems and Computers
Volume2021-October
ISSN (Print)1058-6393

Conference

Conference55th Asilomar Conference on Signals, Systems and Computers, ACSSC 2021
Country/TerritoryUnited States
CityVirtual, Pacific Grove
Period10/31/2111/3/21

Keywords

  • cost-performance tradeoff
  • exploration-exploitation
  • mini-batch
  • stochastic gradient descent

Fingerprint

Dive into the research topics of 'EE-Grad: Exploration and Exploitation for Cost-Efficient Mini-Batch SGD'. Together they form a unique fingerprint.

Cite this