Improving the stability for spiking neural networks using anti-noise learning rule

Yuling Luo, Qiang Fu, Junxiu Liu*, Yongchuang Huang, Xuemei Ding, Yi Cao

*Corresponding author for this work

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

Abstract

Most of the existing SNNs only consider training the noise-free data. However, noise extensively exists in actual SNNs. The stability of networks is affected by noise perturbation during the training period. Therefore, one research challenge is to improve the stability and produce reliable outputs under the present of noises. In this paper, the training method and the exponential method are employed to enhance the neural network ability of noise tolerance. The comparison of conventional and anti-noise SNNs under various tasks shows that the anti-noise SNN can significantly improve the noise tolerance capability.

Original languageEnglish
Title of host publicationPRICAI 2018
Subtitle of host publicationTrends in Artificial Intelligence - 15th Pacific Rim International Conference on Artificial Intelligence, Proceedings
EditorsXin Geng, Byeong-Ho Kang
PublisherSpringer-Verlag
Pages29-37
Number of pages9
ISBN (Print)9783319973098
DOIs
Publication statusPublished - 2018
Event15th Pacific Rim International Conference on Artificial Intelligence, PRICAI 2018 - Nanjing, China
Duration: 28 Aug 201831 Aug 2018

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume11013 LNAI
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference15th Pacific Rim International Conference on Artificial Intelligence, PRICAI 2018
Country/TerritoryChina
CityNanjing
Period28/08/1831/08/18

Keywords

  • learning rule
  • noise tolerance
  • Spiking Neural Network
  • stability

Fingerprint

Dive into the research topics of 'Improving the stability for spiking neural networks using anti-noise learning rule'. Together they form a unique fingerprint.

Cite this