Linear unit relevance in multiclass NLDA networks

Abstract

In this work we shall discuss how to apply classical input relevance results for linear Fisher discriminants to measure the relevance of the linear last hidden layer of a Non Linear Discriminant Analysis (NLDA) network. We shall quickly review first possible ways to extend classical and non linear Fisher analysis to multiclass problems and introduce a criterion function very well suited computationally to NLDA networks. After defining a relevance statistic for linear NLDA units, we shall numerically illustrate the resulting procedures on a synthetic 3 class classification problem. © Springer-Verlag Berlin Heidelberg 2003.