Constructive learning of recurrent neural networks

D. Chen, C. L. Giles, G. Z. Sun, H. H. Chen, Y. C. Lee, M. W. Goudreau

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

7 Scopus citations

Abstract

Recurrent neural networks are a natural model for learning and predicting temporal signals. In addition, simple recurrent networks have been shown to be both theoretically and experimentally capable of learning finite state automata [Cleeremans 89. Giles 92a, Minsky 67, Pollack 91, Siegelmann 92]. However, it is difficult to determine what is the minimal neural network structure for a particular automaton. Using a large recurrent network, which would be versatile in theory, in practice proves to be very difficult to train. Constructive or destructive recurrent methods might offer a solution to this problem. We prove that one current method. Recurrent Cascade Correlation, has fundamental limitations in representation and thus in its learning capabilities. We give a preliminary approach on how to get around these limitations by devising a "simple" constructive training method that adds neurons during training while still preserving the powerful fully recurrent structure. Through simulations we show that such a method can learn many types of regular grammars that the Recurrent Cascade Correlation method is unable to learn.

Original languageEnglish (US)
Title of host publication1993 IEEE International Conference on Neural Networks, ICNN 1993
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages1196-1201
Number of pages6
ISBN (Electronic)0780309995
DOIs
StatePublished - Jan 1 1993
EventIEEE International Conference on Neural Networks, ICNN 1993 - San Francisco, United States
Duration: Mar 28 1993Apr 1 1993

Publication series

NameIEEE International Conference on Neural Networks - Conference Proceedings
Volume1993-January
ISSN (Print)1098-7576

Other

OtherIEEE International Conference on Neural Networks, ICNN 1993
CountryUnited States
CitySan Francisco
Period3/28/934/1/93

All Science Journal Classification (ASJC) codes

  • Software

Fingerprint Dive into the research topics of 'Constructive learning of recurrent neural networks'. Together they form a unique fingerprint.

Cite this