Extending Athena: multiple classes and confidence output values

Abstract

The Athena model, introduced by C. Koutsougeras and C.A. Papachristou (1988), is a tree-like net whose adaptation is based on entropy optimization. The difficult problem in the optimization was handled by using Fisher's linear discriminant method. To handle the multiple class case, heuristics were used (multiple classes, generic classes) to reduce the problem at hand to the two-class case. In the present work, it is shown that the more general Fisher method of multiple discriminants is very effective in directly handling the multiple classes case. A method is also presented by which confidence values are produced for the overall classification decision.<>