A study of various desired response and error scaling sequences for temporal pattern classification using a FIR neural network

Author

Arsenault, N. ; Stevenson, M.

Author_Institution

Dept. of Electr. Eng., New Brunswick Univ., Fredericton, NB, Canada

Volume

1

fYear

1995

Firstpage

194

Abstract

A finite impulse response neural network is configured to act as a temporal pattern classifier. The notion of desired response and error scaling sequences is introduced and the effects of these sequences on the classification rate and network outputs is examined. Narrow error scaling functions speed learning but produce poor quality outputs. Wide error scaling functions produce better quality output but learn more slowly

Keywords

FIR filters; error analysis; filtering theory; learning (artificial intelligence); multilayer perceptrons; pattern classification; sequences; FIR filter; FIR neural network; classification rate; desired response; desired response sequence; error scaling sequence; finite impulse response neural network; learning; narrow error scaling functions; network outputs; temporal pattern classification; wide error scaling functions; Clocks; Feedforward neural networks; Feedforward systems; Finite impulse response filter; Frequency; Neural networks; Pattern classification; Signal processing; Signal to noise ratio; Speech;

fLanguage

English

Publisher

ieee

Conference_Titel

Electrical and Computer Engineering, 1995. Canadian Conference on

Conference_Location

Montreal, Que.

ISSN

0840-7789

Print_ISBN

0-7803-2766-7

Type

conf

DOI

10.1109/CCECE.1995.528107

Filename

528107