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.
CITATION STYLE
Dorronsoro, J. R., González, A., & Serrano, E. (2003). Linear unit relevance in multiclass NLDA networks. Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 2686, 174–181. https://doi.org/10.1007/3-540-44868-3_23
Mendeley helps you to discover research relevant for your work.