### 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. However it is difficult to determine what is the minimal neural network structure for a particular automation. 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 limitation 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 language | English (US) |
---|---|

Title of host publication | 1993 IEEE International Conference on Neural Networks |

Publisher | Publ by IEEE |

Pages | 1192-1201 |

Number of pages | 10 |

ISBN (Print) | 0780312007 |

State | Published - Jan 1 1993 |

Event | 1993 IEEE International Conference on Neural Networks - San Francisco, California, USA Duration: Mar 28 1993 → Apr 1 1993 |

### Publication series

Name | 1993 IEEE International Conference on Neural Networks |
---|

### Other

Other | 1993 IEEE International Conference on Neural Networks |
---|---|

City | San Francisco, California, USA |

Period | 3/28/93 → 4/1/93 |

### Fingerprint

### All Science Journal Classification (ASJC) codes

- Engineering(all)

### Cite this

*1993 IEEE International Conference on Neural Networks*(pp. 1192-1201). (1993 IEEE International Conference on Neural Networks). Publ by IEEE.

}

*1993 IEEE International Conference on Neural Networks.*1993 IEEE International Conference on Neural Networks, Publ by IEEE, pp. 1192-1201, 1993 IEEE International Conference on Neural Networks, San Francisco, California, USA, 3/28/93.

**Constructive learning of recurrent neural network.** / Chen, D.; Giles, C. L.; Sun, G. Z.; Chen, H. H.; Lee, Y. C.; Goudreau, M. W.

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

TY - GEN

T1 - Constructive learning of recurrent neural network

AU - Chen, D.

AU - Giles, C. L.

AU - Sun, G. Z.

AU - Chen, H. H.

AU - Lee, Y. C.

AU - Goudreau, M. W.

PY - 1993/1/1

Y1 - 1993/1/1

N2 - 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. However it is difficult to determine what is the minimal neural network structure for a particular automation. 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 limitation 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.

AB - 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. However it is difficult to determine what is the minimal neural network structure for a particular automation. 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 limitation 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.

UR - http://www.scopus.com/inward/record.url?scp=0027187853&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=0027187853&partnerID=8YFLogxK

M3 - Conference contribution

AN - SCOPUS:0027187853

SN - 0780312007

T3 - 1993 IEEE International Conference on Neural Networks

SP - 1192

EP - 1201

BT - 1993 IEEE International Conference on Neural Networks

PB - Publ by IEEE

ER -