Evaluation of neural networks for automatic target recognition

Author

Przytula, K. Wojtek ; Thompson, Don

Author_Institution

Hughes Res. Labs., Malibu, CA, USA

Volume

3

fYear

1997

fDate

1-8 Feb 1997

Firstpage

423

Abstract

In automatic target recognition we often face a problem in having to train a large neural network upon a very limited data set. This paper presents methods designed to analyze trained networks. The methods allow us to investigate how the network makes its decisions as well as its generalization properties. The methods interact with each other and are intended to be used as a complete set. They use techniques of sensitivity analysis, linear algebra, and rule extraction. They have been coded in Matlab as a toolbox and tested on a large number of real networks

Keywords

covariance matrices; feature extraction; generalisation (artificial intelligence); learning (artificial intelligence); linear algebra; neural nets; object recognition; radar computing; radar target recognition; sensitivity analysis; target tracking; Matlab toolbox; automatic target recognition; covariance matrix; generalization properties; limited data set; linear algebra; neural networks; rule extraction; sensitivity analysis; trained networks analysis; Design methodology; Electronic mail; Infrared sensors; Laboratories; Linear algebra; Mathematics; Neural networks; Sensitivity analysis; Target recognition; Testing;

fLanguage

English

Publisher

ieee

Conference_Titel

Aerospace Conference, 1997. Proceedings., IEEE

Conference_Location

Snowmass at Aspen, CO

Print_ISBN

0-7803-3741-7

Type

conf

DOI

10.1109/AERO.1997.574896

Filename

574896