Convolutional neural networks considering robustness improvement and its application to face recognition

Amin Jalali, Giljin Jang, Jun Su Kang, Minho Lee

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

10 Scopus citations

Abstract

This paper proposes a novel activation function to promote robustness to the outliers of the training samples. Data samples in the decision boundaries are weighted more by adding the derivatives of the sigmoid function outputs to avoid drastic update of the network weights. Therefore, the network becomes more robust to outliers and noisy patterns. We also present appropriate backpropagation learning algorithm for the convolutional neural networks. We evaluate the performance improvement by the proposed method on a face recognition task, and proved that it outperformed the state of art face recognition methods.

Original languageEnglish
Title of host publicationNeural Information Processing - 22nd International Conference, ICONIP 2015, Proceedings
EditorsSabri Arik, Tingwen Huang, Weng Kin Lai, Qingshan Liu
PublisherSpringer Verlag
Pages240-245
Number of pages6
ISBN (Print)9783319265605
DOIs
StatePublished - 2015
Event22nd International Conference on Neural Information Processing, ICONIP 2015 - Istanbul, Turkey
Duration: 9 Nov 201512 Nov 2015

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume9492
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference22nd International Conference on Neural Information Processing, ICONIP 2015
Country/TerritoryTurkey
CityIstanbul
Period9/11/1512/11/15

Keywords

  • Back propagation
  • Convolutional neural network
  • Deep learning
  • Gradient descent
  • Robustness in cost function

Fingerprint

Dive into the research topics of 'Convolutional neural networks considering robustness improvement and its application to face recognition'. Together they form a unique fingerprint.

Cite this