Evidence Based Reasoning in Classifier Hierarchies

Rebecca Fay, Friedhelm Schwenker, Günther Palm

Hierarchical neural networks naturally combine subsymbolic information processing with symbolic information as they consist of several neural classifiers which provide hierarchically structured knowledge. This knowledge implies a particular uncertainty which is indicated by the magnitude of the classifier outputs. There are different ways to combine this expert knowledge to a collective output. Two different methods are evaluated in this paper: a method similar to the decision tree approach and an evidence theoretic approach utilising Dempster-Shafer theory. The proposed approaches have been evaluated using three different data sets and two different types of classifiers. It was shown that the evidence theoretic approach yields improved classification performance.