CNNs with Compact Activation Function

Jindong Wang, Jinchao Xu, Jianqing Zhu

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

3 Scopus citations

Abstract

Activation function plays an important role in neural networks. We propose to use hat activation function, namely the first order B-spline, as activation function for CNNs including MgNet and ResNet. Different from commonly used activation functions like ReLU, the hat function has a compact support and no obvious spectral bias. Although spectral bias is thought to be beneficial for generalization, we show that MgNet and ResNet with hat function still exhibit a slightly better generalization performance than CNNs with ReLU function by our experiments of classification on MNIST, CIFAR10/100 and ImageNet datasets. This indicates that CNNs without spectral bias can have a good generalization capability. We also illustrate that although hat function has a small activation area which is more likely to induce vanishing gradient problem, hat CNNs with various initialization methods still works well.

Original languageEnglish (US)
Title of host publicationComputational Science - ICCS 2022, 22nd International Conference, Proceedings
EditorsDerek Groen, Clélia de Mulatier, Valeria V. Krzhizhanovskaya, Peter M.A. Sloot, Maciej Paszynski, Jack J. Dongarra
PublisherSpringer Science and Business Media Deutschland GmbH
Pages319-327
Number of pages9
ISBN (Print)9783031087530
DOIs
StatePublished - 2022
Event22nd Annual International Conference on Computational Science, ICCS 2022 - London, United Kingdom
Duration: Jun 21 2022Jun 23 2022

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume13351 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference22nd Annual International Conference on Computational Science, ICCS 2022
Country/TerritoryUnited Kingdom
CityLondon
Period6/21/226/23/22

All Science Journal Classification (ASJC) codes

  • Theoretical Computer Science
  • General Computer Science

Fingerprint

Dive into the research topics of 'CNNs with Compact Activation Function'. Together they form a unique fingerprint.

Cite this