PT Chapter
AU Fernando Vilariño
   Debora Gil
   Petia Radeva
TI A Novel FLDA Formulation for Numerical Stability Analysis
BT Recent Advances in Artificial Intelligence Research and Development
PY 2004
BP 77
EP 84
VL 113
DE Supervised Learning; Linear Discriminant Analysis; Numerical Stability; Computer Vision
AB Fisher Linear Discriminant Analysis (FLDA) is one of the most popular techniques used in classification applying dimensional reduction. The numerical scheme involves the inversion of the within-class scatter matrix, which makes FLDA potentially ill-conditioned when it becomes singular. In this paper we present a novel explicit formulation of FLDA in terms of the eccentricity ratio and eigenvector orientations of the within-class scatter matrix. An analysis of this function will characterize those situations where FLDA response is not reliable because of numerical instability. This can solve common situations of poor classification performance in computer vision.
ER