Using ancillary statistics in on-line learning algorithms

Huaiyu Zhu; Richard Rohwer

Using ancillary statistics in on-line learning algorithms

Huaiyu Zhu, Richard Rohwer

Research output: Chapter in Book/Published conference output › Chapter

Abstract

Neural networks are usually curved statistical models. They do not have finite dimensional sufficient statistics, so on-line learning on the model itself inevitably loses information. In this paper we propose a new scheme for training curved models, inspired by the ideas of ancillary statistics and adaptive critics. At each point estimate an auxiliary flat model (exponential family) is built to locally accommodate both the usual statistic (tangent to the model) and an ancillary statistic (normal to the model). The auxiliary model plays a role in determining credit assignment analogous to that played by an adaptive critic in solving temporal problems. The method is illustrated with the Cauchy model and the algorithm is proved to be asymptotically efficient.

Original language	English
Title of host publication	Proceedings of the 1996 International Conference on Neural Information Processing
Editors	S I Amari, L Xu, L W Chan, I King, K S Leung
Publisher	Springer
ISBN (Print)	9789813083059
Publication status	Published - 1996
Event	Proc. 1996 International Conference on Neural Information Processing - Duration: 1 Jan 1996 → 1 Jan 1996

Conference

Conference	Proc. 1996 International Conference on Neural Information Processing
Period	1/01/96 → 1/01/96

Bibliographical note

The original publication is available at www.springerlink.com

Keywords

Neural networks
curved models
auxiliary
ancillary statistic
Cauchy
algorithm

Access to Document

Using ancillary statistics in on-line learning algorithms
The original publication is available at www.springerlink.com
Accepted author manuscript, 228 KB

Cite this

@inbook{a729e56db2fb45b485b99acd19179c8a,

title = "Using ancillary statistics in on-line learning algorithms",

abstract = "Neural networks are usually curved statistical models. They do not have finite dimensional sufficient statistics, so on-line learning on the model itself inevitably loses information. In this paper we propose a new scheme for training curved models, inspired by the ideas of ancillary statistics and adaptive critics. At each point estimate an auxiliary flat model (exponential family) is built to locally accommodate both the usual statistic (tangent to the model) and an ancillary statistic (normal to the model). The auxiliary model plays a role in determining credit assignment analogous to that played by an adaptive critic in solving temporal problems. The method is illustrated with the Cauchy model and the algorithm is proved to be asymptotically efficient.",

keywords = "Neural networks, curved models, auxiliary, ancillary statistic, Cauchy, algorithm",

author = "Huaiyu Zhu and Richard Rohwer",

note = "The original publication is available at www.springerlink.com; Proc. 1996 International Conference on Neural Information Processing ; Conference date: 01-01-1996 Through 01-01-1996",

year = "1996",

language = "English",

isbn = "9789813083059",

editor = "Amari, {S I} and L Xu and Chan, {L W} and I King and Leung, {K S}",

booktitle = "Proceedings of the 1996 International Conference on Neural Information Processing",

publisher = "Springer",

address = "Germany",

}

TY - CHAP

T1 - Using ancillary statistics in on-line learning algorithms

AU - Zhu, Huaiyu

AU - Rohwer, Richard

N1 - The original publication is available at www.springerlink.com

PY - 1996

Y1 - 1996

N2 - Neural networks are usually curved statistical models. They do not have finite dimensional sufficient statistics, so on-line learning on the model itself inevitably loses information. In this paper we propose a new scheme for training curved models, inspired by the ideas of ancillary statistics and adaptive critics. At each point estimate an auxiliary flat model (exponential family) is built to locally accommodate both the usual statistic (tangent to the model) and an ancillary statistic (normal to the model). The auxiliary model plays a role in determining credit assignment analogous to that played by an adaptive critic in solving temporal problems. The method is illustrated with the Cauchy model and the algorithm is proved to be asymptotically efficient.

AB - Neural networks are usually curved statistical models. They do not have finite dimensional sufficient statistics, so on-line learning on the model itself inevitably loses information. In this paper we propose a new scheme for training curved models, inspired by the ideas of ancillary statistics and adaptive critics. At each point estimate an auxiliary flat model (exponential family) is built to locally accommodate both the usual statistic (tangent to the model) and an ancillary statistic (normal to the model). The auxiliary model plays a role in determining credit assignment analogous to that played by an adaptive critic in solving temporal problems. The method is illustrated with the Cauchy model and the algorithm is proved to be asymptotically efficient.

KW - Neural networks

KW - curved models

KW - auxiliary

KW - ancillary statistic

KW - Cauchy

KW - algorithm

UR - http://www.springer.com/computer/artificial/book/978-981-3083-05-9

M3 - Chapter

SN - 9789813083059

BT - Proceedings of the 1996 International Conference on Neural Information Processing

A2 - Amari, S I

A2 - Xu, L

A2 - Chan, L W

A2 - King, I

A2 - Leung, K S

PB - Springer

T2 - Proc. 1996 International Conference on Neural Information Processing

Y2 - 1 January 1996 through 1 January 1996

ER -

Using ancillary statistics in on-line learning algorithms

Abstract

Conference

Bibliographical note

Keywords

Access to Document

Other files and links

Fingerprint

Cite this