Metamorphosis Networks: An Alternative to Constructive Methods

Brian V. Bonnlander; Michael C. Mozer

doi:10.5555/645753.668052

Metamorphosis Networks: An Alternative to Constructive Methods

Brian V. Bonnlander
, Michael C. Mozer

University of Colorado Boulder

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

1 Scopus citations

Abstract

Given a set of training examples, determining the appropriate number of free parameters is a challenging problem. Constructive learning algorithms attempt to solve this problem automatically by adding hidden units, and therefore free parameters, during learning. We explore an alternative class of algorithms-called metamorphosis algorithms-in which the number of units is fixed, but the number of free parameters gradually increases during learning. The architecture we investigate is composed of RBF units on a lattice, which imposes flexible constraints on the parameters of the network. Virtues of this approach include variable subset selection, robust parameter selection, multiresolution processing, and interpolation of sparse training data.

Original language	English
Title of host publication	Advances in Neural Information Processing Systems 5, NIPS 1992
Editors	Stephen Jose Hanson, Jack D. Cowan, C. Lee Giles
Publisher	Neural information processing systems foundation
Pages	131-138
Number of pages	8
ISBN (Electronic)	1558602747, 9781558602748
DOIs	https://doi.org/10.5555/645753.668052
State	Published - 1992
Externally published	Yes
Event	5th Advances in Neural Information Processing Systems, NIPS 1992 - Denver, United States Duration: Nov 30 1992 → Dec 3 1992

Publication series

Name	Advances in Neural Information Processing Systems
Volume	5
ISSN (Print)	1049-5258

Conference

Conference	5th Advances in Neural Information Processing Systems, NIPS 1992
Country/Territory	United States
City	Denver
Period	11/30/92 → 12/3/92

Access to Document

10.5555/645753.668052

Cite this

Bonnlander, B. V., & Mozer, M. C. (1992). Metamorphosis Networks: An Alternative to Constructive Methods. In S. J. Hanson, J. D. Cowan, & C. L. Giles (Eds.), Advances in Neural Information Processing Systems 5, NIPS 1992 (pp. 131-138). (Advances in Neural Information Processing Systems; Vol. 5). Neural information processing systems foundation. https://doi.org/10.5555/645753.668052

@inproceedings{219020ff4d7f4d6ea35d3e8faf4a9a3f,

title = "Metamorphosis Networks: An Alternative to Constructive Methods",

abstract = "Given a set of training examples, determining the appropriate number of free parameters is a challenging problem. Constructive learning algorithms attempt to solve this problem automatically by adding hidden units, and therefore free parameters, during learning. We explore an alternative class of algorithms-called metamorphosis algorithms-in which the number of units is fixed, but the number of free parameters gradually increases during learning. The architecture we investigate is composed of RBF units on a lattice, which imposes flexible constraints on the parameters of the network. Virtues of this approach include variable subset selection, robust parameter selection, multiresolution processing, and interpolation of sparse training data.",

author = "Bonnlander, \{Brian V.\} and Mozer, \{Michael C.\}",

note = "Publisher Copyright: {\textcopyright} 1992 Neural information processing systems foundation. All rights reserved.; 5th Advances in Neural Information Processing Systems, NIPS 1992 ; Conference date: 30-11-1992 Through 03-12-1992",

year = "1992",

doi = "10.5555/645753.668052",

language = "English",

series = "Advances in Neural Information Processing Systems",

publisher = "Neural information processing systems foundation",

pages = "131--138",

editor = "Hanson, \{Stephen Jose\} and Cowan, \{Jack D.\} and Giles, \{C. Lee\}",

booktitle = "Advances in Neural Information Processing Systems 5, NIPS 1992",

address = "United States",

}

Bonnlander, BV & Mozer, MC 1992, Metamorphosis Networks: An Alternative to Constructive Methods. in SJ Hanson, JD Cowan & CL Giles (eds), Advances in Neural Information Processing Systems 5, NIPS 1992. Advances in Neural Information Processing Systems, vol. 5, Neural information processing systems foundation, pp. 131-138, 5th Advances in Neural Information Processing Systems, NIPS 1992, Denver, United States, 11/30/92. https://doi.org/10.5555/645753.668052

Metamorphosis Networks: An Alternative to Constructive Methods. / Bonnlander, Brian V.; Mozer, Michael C.
Advances in Neural Information Processing Systems 5, NIPS 1992. ed. / Stephen Jose Hanson; Jack D. Cowan; C. Lee Giles. Neural information processing systems foundation, 1992. p. 131-138 (Advances in Neural Information Processing Systems; Vol. 5).

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

TY - GEN

T1 - Metamorphosis Networks

T2 - 5th Advances in Neural Information Processing Systems, NIPS 1992

AU - Bonnlander, Brian V.

AU - Mozer, Michael C.

PY - 1992

Y1 - 1992

N2 - Given a set of training examples, determining the appropriate number of free parameters is a challenging problem. Constructive learning algorithms attempt to solve this problem automatically by adding hidden units, and therefore free parameters, during learning. We explore an alternative class of algorithms-called metamorphosis algorithms-in which the number of units is fixed, but the number of free parameters gradually increases during learning. The architecture we investigate is composed of RBF units on a lattice, which imposes flexible constraints on the parameters of the network. Virtues of this approach include variable subset selection, robust parameter selection, multiresolution processing, and interpolation of sparse training data.

AB - Given a set of training examples, determining the appropriate number of free parameters is a challenging problem. Constructive learning algorithms attempt to solve this problem automatically by adding hidden units, and therefore free parameters, during learning. We explore an alternative class of algorithms-called metamorphosis algorithms-in which the number of units is fixed, but the number of free parameters gradually increases during learning. The architecture we investigate is composed of RBF units on a lattice, which imposes flexible constraints on the parameters of the network. Virtues of this approach include variable subset selection, robust parameter selection, multiresolution processing, and interpolation of sparse training data.

UR - https://www.scopus.com/pages/publications/105021000529

U2 - 10.5555/645753.668052

DO - 10.5555/645753.668052

M3 - Conference contribution

AN - SCOPUS:105021000529

T3 - Advances in Neural Information Processing Systems

SP - 131

EP - 138

BT - Advances in Neural Information Processing Systems 5, NIPS 1992

A2 - Hanson, Stephen Jose

A2 - Cowan, Jack D.

A2 - Giles, C. Lee

PB - Neural information processing systems foundation

Y2 - 30 November 1992 through 3 December 1992

ER -

Metamorphosis Networks: An Alternative to Constructive Methods

Abstract

Publication series

Conference

Access to Document

Other files and links

Fingerprint

Cite this