No more pesky learning rates

Tom Schaul, Sixin Zhang, Yann LeCun

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

The performance of stochastic gradient descent (SGD) depends critically on how learning rates are tuned and decreased over time. We propose a method to automatically adjust multiple learning rates so as to minimize the expected error at any one time. The method relies on local gradient variations across samples. In our approach, learning rates can increase as well as decrease, making it suitable for non-stationary problems. Using a number of convex and non-convex learning tasks, we show that the resulting algorithm matches the performance of SGD or other adaptive approaches with their best settings obtained through systematic search, and effectively removes the need for learning rate tuning.

Original languageEnglish (US)
Title of host publication30th International Conference on Machine Learning, ICML 2013
PublisherInternational Machine Learning Society (IMLS)
Pages1380-1388
Number of pages9
EditionPART 2
StatePublished - 2013
Event30th International Conference on Machine Learning, ICML 2013 - Atlanta, GA, United States
Duration: Jun 16 2013Jun 21 2013

Other

Other30th International Conference on Machine Learning, ICML 2013
CountryUnited States
CityAtlanta, GA
Period6/16/136/21/13

Fingerprint

Tuning
learning
performance
time

ASJC Scopus subject areas

  • Human-Computer Interaction
  • Sociology and Political Science

Cite this

Schaul, T., Zhang, S., & LeCun, Y. (2013). No more pesky learning rates. In 30th International Conference on Machine Learning, ICML 2013 (PART 2 ed., pp. 1380-1388). International Machine Learning Society (IMLS).

No more pesky learning rates. / Schaul, Tom; Zhang, Sixin; LeCun, Yann.

30th International Conference on Machine Learning, ICML 2013. PART 2. ed. International Machine Learning Society (IMLS), 2013. p. 1380-1388.

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Schaul, T, Zhang, S & LeCun, Y 2013, No more pesky learning rates. in 30th International Conference on Machine Learning, ICML 2013. PART 2 edn, International Machine Learning Society (IMLS), pp. 1380-1388, 30th International Conference on Machine Learning, ICML 2013, Atlanta, GA, United States, 6/16/13.
Schaul T, Zhang S, LeCun Y. No more pesky learning rates. In 30th International Conference on Machine Learning, ICML 2013. PART 2 ed. International Machine Learning Society (IMLS). 2013. p. 1380-1388
Schaul, Tom ; Zhang, Sixin ; LeCun, Yann. / No more pesky learning rates. 30th International Conference on Machine Learning, ICML 2013. PART 2. ed. International Machine Learning Society (IMLS), 2013. pp. 1380-1388
@inproceedings{eafa72d132c64e4eb285a76496dc3c7f,
title = "No more pesky learning rates",
abstract = "The performance of stochastic gradient descent (SGD) depends critically on how learning rates are tuned and decreased over time. We propose a method to automatically adjust multiple learning rates so as to minimize the expected error at any one time. The method relies on local gradient variations across samples. In our approach, learning rates can increase as well as decrease, making it suitable for non-stationary problems. Using a number of convex and non-convex learning tasks, we show that the resulting algorithm matches the performance of SGD or other adaptive approaches with their best settings obtained through systematic search, and effectively removes the need for learning rate tuning.",
author = "Tom Schaul and Sixin Zhang and Yann LeCun",
year = "2013",
language = "English (US)",
pages = "1380--1388",
booktitle = "30th International Conference on Machine Learning, ICML 2013",
publisher = "International Machine Learning Society (IMLS)",
edition = "PART 2",

}

TY - GEN

T1 - No more pesky learning rates

AU - Schaul, Tom

AU - Zhang, Sixin

AU - LeCun, Yann

PY - 2013

Y1 - 2013

N2 - The performance of stochastic gradient descent (SGD) depends critically on how learning rates are tuned and decreased over time. We propose a method to automatically adjust multiple learning rates so as to minimize the expected error at any one time. The method relies on local gradient variations across samples. In our approach, learning rates can increase as well as decrease, making it suitable for non-stationary problems. Using a number of convex and non-convex learning tasks, we show that the resulting algorithm matches the performance of SGD or other adaptive approaches with their best settings obtained through systematic search, and effectively removes the need for learning rate tuning.

AB - The performance of stochastic gradient descent (SGD) depends critically on how learning rates are tuned and decreased over time. We propose a method to automatically adjust multiple learning rates so as to minimize the expected error at any one time. The method relies on local gradient variations across samples. In our approach, learning rates can increase as well as decrease, making it suitable for non-stationary problems. Using a number of convex and non-convex learning tasks, we show that the resulting algorithm matches the performance of SGD or other adaptive approaches with their best settings obtained through systematic search, and effectively removes the need for learning rate tuning.

UR - http://www.scopus.com/inward/record.url?scp=84897487847&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=84897487847&partnerID=8YFLogxK

M3 - Conference contribution

AN - SCOPUS:84897487847

SP - 1380

EP - 1388

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

PB - International Machine Learning Society (IMLS)

ER -