Non-Linear Variance Reduction Techniques in Biometric Authentication
Norman Poh, Samy Bengio · Infoscience (Ecole Polytechnique Fédérale de Lausanne) · 2003
Abstract — In this paper, several approaches that can be used to improve biometric authentication applications are proposed. The idea is inspired by the ensemble approach, i.e., the use of several classifiers to solve a problem. Compared to using only one classifier, the ensemble of classifiers has the advantage of reducing the overall variance of the system. Instead of using multiple classifiers, we propose here to examine other possible means of variance reduction (VR), namely through the use of multiple synthetic samples, different extractors (features) and biometric modalities. The scores are combined using the average operator, Multi-Layer Perceptrons and Support Vector Machines. It is found empirically that VR via modalities is the best technique, followed by VR via extractors, VR via classifiers and VR via synthetic samples. This decreasing order of effectiveness is due to the corresponding degree of independence of the combined objects. The theoretical and empirical findings show that experts combined via VR techniques always perform better than the average of their participating experts. Furthermore, in practice, most combined experts perform better than any of their participating experts. I.