Mutual features for pattern classification

FACULTY OF ENGINEERING, SCIENCE AND MATHEMATICS SCHOOL OF ELECTRONICS AND COMPUTER SCIENCE Doctor of Philosophy MUTUAL FEATURES FOR PATTERN CLASSIFICATION by Heiko Claussen The mean of a data set is one trivial representation of data from one class. This thesis discusses mutual interdependence analysis (MIA) that is successfully used to extract more involved representations, or “mutual features”, accounting for samples in the class. MIA aims to extract a common or mutual signature that is invariant to changes in the inputs. For example, a mutual feature is a speaker signature under varying channel conditions or a face signature under varying illumination conditions. By definition, the mutual feature is a linear combination of class examples that is equally correlated with all training samples in the class. An equivalent view is to find a direction to project the dataset such that projection lengths are maximally correlated. The MIA optimization criterion is presented from the perspectives of canonical correlation analysis and Bayesian estimation. This allows to state and solve the criterion for mutual features concisely and to infer other properties of its closed form, unique solution under various statistical assumptions. Moreover, a generalized MIA solution (GMIA) is defined that enables utilization of a priori knowledge. MIA and GMIA work well even if the mutual signature accounts only for a small part of the energy in the inputs. Real world problems do not exactly fit the signal model of an equally correlated common signature. Therefore, the behavior of MIA is analyzed in situations where its model does not exactly fit. For these situations it is shown that GMIA continues to extract meaningful information. Furthermore, the GMIA result is compared to ubiquitous signal processing methods. It is shown that GMIA extends these current tools visualizing previously hidden information. The utility of both MIA and GMIA is demonstrated on two standard pattern recognition problems: text–independent speaker verification and illumination–independent face recognition. For example, GMIA achieves an equal error rate (EER) of 4.0% in the text–independent speaker verification application on the full NTIMIT database of 630 speakers. On the other hand, for illumination–independent face recognition, MIA achieves an identification error rate of 7.4% in exhausive leave–one–out tests on the Yale database. Overall, MIA and GMIA are found to achieve competitive pattern classification performance to other modern algorithms.