### Abstract

Original language | English |
---|---|

Title of host publication | On-line learning in neural networks |

Editors | David Saad |

Place of Publication | Isaac Newton Institute, Cambridge |

Publisher | Cambridge University Press |

Pages | 183-207 |

Number of pages | 25 |

ISBN (Print) | 0521652634 |

DOIs | |

Publication status | Published - Jan 1999 |

Event | Proceedings of the on-line learning themed week - Duration: 1 Jan 1999 → 1 Jan 1999 |

### Publication series

Name | Publications of the Newton Institute |
---|---|

Publisher | Cambridge University Press |

### Other

Other | Proceedings of the on-line learning themed week |
---|---|

Period | 1/01/99 → 1/01/99 |

### Fingerprint

### Bibliographical note

Copyright of the Institute of Electrical and Electronics Engineers (IEEE)### Keywords

- multi-layer neural networks
- statistical mechanics
- optimal asymptotic performance
- asymptotic generalization error decay
- Matrix momentum

### Cite this

*On-line learning in neural networks*(pp. 183-207). (Publications of the Newton Institute). Isaac Newton Institute, Cambridge: Cambridge University Press. https://doi.org/10.2277/0521652634

}

*On-line learning in neural networks.*Publications of the Newton Institute, Cambridge University Press, Isaac Newton Institute, Cambridge, pp. 183-207, Proceedings of the on-line learning themed week, 1/01/99. https://doi.org/10.2277/0521652634

**Incorporating curvature information into on-line learning.** / Rattray, Magnus; Saad, David.

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

TY - CHAP

T1 - Incorporating curvature information into on-line learning

AU - Rattray, Magnus

AU - Saad, David

N1 - Copyright of the Institute of Electrical and Electronics Engineers (IEEE)

PY - 1999/1

Y1 - 1999/1

N2 - We analyse the dynamics of a number of second order on-line learning algorithms training multi-layer neural networks, using the methods of statistical mechanics. We first consider on-line Newton's method, which is known to provide optimal asymptotic performance. We determine the asymptotic generalization error decay for a soft committee machine, which is shown to compare favourably with the result for standard gradient descent. Matrix momentum provides a practical approximation to this method by allowing an efficient inversion of the Hessian. We consider an idealized matrix momentum algorithm which requires access to the Hessian and find close correspondence with the dynamics of on-line Newton's method. In practice, the Hessian will not be known on-line and we therefore consider matrix momentum using a single example approximation to the Hessian. In this case good asymptotic performance may still be achieved, but the algorithm is now sensitive to parameter choice because of noise in the Hessian estimate. On-line Newton's method is not appropriate during the transient learning phase, since a suboptimal unstable fixed point of the gradient descent dynamics becomes stable for this algorithm. A principled alternative is to use Amari's natural gradient learning algorithm and we show how this method provides a significant reduction in learning time when compared to gradient descent, while retaining the asymptotic performance of on-line Newton's method.

AB - We analyse the dynamics of a number of second order on-line learning algorithms training multi-layer neural networks, using the methods of statistical mechanics. We first consider on-line Newton's method, which is known to provide optimal asymptotic performance. We determine the asymptotic generalization error decay for a soft committee machine, which is shown to compare favourably with the result for standard gradient descent. Matrix momentum provides a practical approximation to this method by allowing an efficient inversion of the Hessian. We consider an idealized matrix momentum algorithm which requires access to the Hessian and find close correspondence with the dynamics of on-line Newton's method. In practice, the Hessian will not be known on-line and we therefore consider matrix momentum using a single example approximation to the Hessian. In this case good asymptotic performance may still be achieved, but the algorithm is now sensitive to parameter choice because of noise in the Hessian estimate. On-line Newton's method is not appropriate during the transient learning phase, since a suboptimal unstable fixed point of the gradient descent dynamics becomes stable for this algorithm. A principled alternative is to use Amari's natural gradient learning algorithm and we show how this method provides a significant reduction in learning time when compared to gradient descent, while retaining the asymptotic performance of on-line Newton's method.

KW - multi-layer neural networks

KW - statistical mechanics

KW - optimal asymptotic performance

KW - asymptotic generalization error decay

KW - Matrix momentum

UR - http://www.cambridge.org/uk/catalogue/catalogue.asp?isbn=0521652634

U2 - 10.2277/0521652634

DO - 10.2277/0521652634

M3 - Chapter

SN - 0521652634

T3 - Publications of the Newton Institute

SP - 183

EP - 207

BT - On-line learning in neural networks

A2 - Saad, David

PB - Cambridge University Press

CY - Isaac Newton Institute, Cambridge

ER -