Measurement criteria for neural network pruning

Author

Erdogan, Sevki S. ; Ng, Geok-See ; Chan, P.K.-H.

Author_Institution

Sch. of Appl. Sci., Nanyang Technol. Univ., Singapore

Volume

1

fYear

1996

fDate

26-29 Nov 1996

Firstpage

83

Abstract

A new measure based on hidden-output node activation is proposed for measuring the relevance of hidden nodes in a neural network. The concept has been successfully applied for pruning in several classification problems. The experiments indicate that redundant nodes are pruned down resulting in optimal network topologies. The measure has been compared to the one proposed by Kamimura-Nakanishi (see IEICE Trans. Inf. & Syst., vol.E78-D, no.4, p.484-9, 1995) and also used in the context of a modified cost function where an additional penalty function is added to steer the direction of the hidden node´s activation in the process of learning

Keywords

entropy; learning (artificial intelligence); network topology; neural nets; classification problems; entropy pruning; experiments; hidden-output node activation; learning; measurement criteria; modified cost function; neural network pruning; optimal network topologies; penalty function; Biological neural networks; Cost function; Entropy; Error probability; Network topology; Neural networks; Neurons; Optimization methods; Proposals; Sensitivity analysis;

fLanguage

English

Publisher

ieee

Conference_Titel

TENCON '96. Proceedings., 1996 IEEE TENCON. Digital Signal Processing Applications

Conference_Location

Perth, WA

Print_ISBN

0-7803-3679-8

Type

conf

DOI

10.1109/TENCON.1996.608715

Filename

608715