Skip to main navigation Skip to search Skip to main content

Analysis on Dropout Regularization

Research output: Chapters, Conference Papers, Creative and Literary WorksRGC 12 - Chapter in an edited book (Author)peer-review

Abstract

Dropout, including Bernoulli dropout (equivalently random node fault) and multiplicative Gaussian noise (MGN) dropout (equivalently multiplicative node noise), has been a technique in training a neural network (NN) to achieve better performance. While simulation results have demonstrated its success, not many work has been done to explain why it works (or why it does not work). In this paper, the objective functions (w) of the learning algorithms with Bernoulli dropout and MGN dropout are derived and thus their regularization effects are analyzed. It is found that learning with Bernoulli dropout cannot improve the generalization of a NN if its weights are not scaled down after training. If we further let J (w) be the desired measure of a NN with such inherent dropout, we clarify a misconception that (w) = J (w). The model attained by learning with dropout is not the desired model that can tolerate such inherent dropout.
Original languageEnglish
Title of host publicationNeural Information Processing - 26th International Conference, ICONIP 2019, Proceedings
EditorsTom Gedeon, Kok Wai Wong, Minho Lee
PublisherSpringer 
Pages253-261
ISBN (Electronic)9783030368029
ISBN (Print)9783030368012
DOIs
Publication statusPublished - 2019
Event26th International Conference on Neural Information Processing (ICONIP 2019) - Sydney, Australia
Duration: 12 Dec 201915 Dec 2019

Publication series

NameCommunications in Computer and Information Science
Volume1143
ISSN (Print)1865-0929
ISSN (Electronic)1865-0937

Conference

Conference26th International Conference on Neural Information Processing (ICONIP 2019)
Abbreviated titleAPNNS 2019
PlaceAustralia
CitySydney
Period12/12/1915/12/19

Research Keywords

  • Dropout
  • Multiplicative node noise
  • Random node fault

Fingerprint

Dive into the research topics of 'Analysis on Dropout Regularization'. Together they form a unique fingerprint.

Cite this