Modular neural network architecture using piece-wise linear mapping

Author

Subbarayan, Saravanan ; Kim, Kyung K. ; Manry, Michael T. ; Devarajan, Venkat ; Chen, Hung-Han

Author_Institution

Dept. of Electr. Eng., Texas Univ., Arlington, TX, USA

fYear

1996

fDate

3-6 Nov. 1996

Firstpage

1171

Abstract

A new modular neural network for functional mapping is presented. A training algorithm for the network is presented which employs a clustering method, a weighted distance measure, and the deign of simple modules. Since the individual modules are linear, the network implements a piece-wise linear mapping. The efficiency of this structure in terms of training time and pattern storage capacity is discussed and the results of comparative performances with the multilayer preceptron, is presented. Examples are provided to verify the properties of the modular network.

Keywords

approximation theory; learning (artificial intelligence); modules; multilayer perceptrons; neural net architecture; piecewise-linear techniques; clustering method; efficiency; functional approximation; functional mapping; linear modules; modular network; modular neural network architecture; module design; multilayer preceptron; pattern storage capacity; piecewise linear mapping; training algorithm; training time; weighted distance measure; Aircraft propulsion; Clustering algorithms; Clustering methods; Multi-layer neural network; Neural networks; Piecewise linear techniques; Postal services; Unsupervised learning; Vectors; Weight measurement;

fLanguage

English

Publisher

ieee

Conference_Titel

Signals, Systems and Computers, 1996. Conference Record of the Thirtieth Asilomar Conference on

Conference_Location

Pacific Grove, CA, USA

ISSN

1058-6393

Print_ISBN

0-8186-7646-9

Type

conf

DOI

10.1109/ACSSC.1996.599129

Filename

599129