How to “secure” the decisions of a NN classifier

Author

Decaestecker, Christine ; Van de Merckt, Thierry

Author_Institution

IRIDIA, Univ. Libre de Bruxelles, Belgium

Volume

1

fYear

1994

fDate

27 Jun-2 Jul 1994

Firstpage

263

Abstract

The paper presents an approach to give an introspective capacity to a neural net (NN) classifier. More precisely, a strategy is developed to detect “dangerous” areas in the pattern space where the decisions can be erroneous. In these areas the classification is hazardous and it is preferable not to take a decision. Our approach is detailed for a NN using prototypes named NNP and is based on a geometrical interpretation of the concept representations generated by the NN classifier. Experiments show the advantages of this approach in presence of nonlinear class boundaries, class overlapping and noise

Keywords

computational geometry; decision theory; feedforward neural nets; pattern classification; NN classifier; NNP; class overlapping; concept representations; geometrical interpretation; introspective capacity; neural net classifier; nonlinear class boundaries; pattern space; three-layer fully connected feedforward net; Bayesian methods; Euclidean distance; Multilayer perceptrons; Neural networks; Partitioning algorithms; Prototypes; Radial basis function networks; Temperature dependence; Training data; Zirconium;

fLanguage

English

Publisher

ieee

Conference_Titel

Neural Networks, 1994. IEEE World Congress on Computational Intelligence., 1994 IEEE International Conference on

Conference_Location

Orlando, FL

Print_ISBN

0-7803-1901-X

Type

conf

DOI

10.1109/ICNN.1994.374172

Filename

374172