Reduction of adjusting weights space dimension in feedforward artificial neural networks training

Author

Blyumin, Sam L. ; Saraev, Paul V.

Author_Institution

Dept. of Appl. Math., Lipetsk State Tech. Univ., Russia

fYear

2002

fDate

2002

Firstpage

242

Lastpage

247

Abstract

This report provides an approach to the reduction of the adjusting weights space dimension in two-layer multioutput feedforward artificial neural networks training. Our approach is based on linear-nonlinear network structure with respect to weights. Two training algorithms based on the Newton and Gauss method with pseudo-inversion for optimization were deduced. Training algorithms are extended to multilayer networks. The report carries the information about the analysis of the proposed training algorithms. Results of numerical experiments are also included.

Keywords

feedforward neural nets; learning (artificial intelligence); matrix algebra; multilayer perceptrons; optimisation; Gauss method; Newton method; adjusting weights space dimension; feedforward neural networks; linear-nonlinear network structure; matrices; multilayer networks; neural training; numerical experiments; optimization; pseudo-inversion; two-layer multioutput neural networks; Artificial neural networks; Electronic mail; Encoding; Gaussian processes; Intelligent networks; Mathematics; Multi-layer neural network; Neural networks; Neurons; Optimization methods;

fLanguage

English

Publisher

ieee

Conference_Titel

Artificial Intelligence Systems, 2002. (ICAIS 2002). 2002 IEEE International Conference on

Print_ISBN

0-7695-1733-1

Type

conf

DOI

10.1109/ICAIS.2002.1048097

Filename

1048097