Subspace and graph methods to leverage auxiliary data for limited target data multi-class classification, applied to speaker verification

Abstract

Multi-class classification can be adversely affected by the absence of sufficient target (in-class) instances for training. Such cases arise in face recognition, speaker verification, and document classification, among others. Auxiliary data-sets, which contain a diverse sampling of non-target instances, are leveraged in this thesis using subspace and graph methods to improve classification where target data is limited. The auxiliary data is used to define a compact representation that maps instances into a vector space where inner products quantify class similarity. Within this space, an estimate of the subspace that constitutes within-class variability (e.g. the recording channel in speaker verification or the illumination conditions in face recognition) can be obtained using class-labeled auxiliary data. This thesis proposes a way to incorporate this estimate into the SVM framework to perform nuisance compensation, thus improving classification performance. Another contribution is a framework that combines mapping and compensation into a single linear comparison, which motivates computationally inexpensive and accurate comparison functions. A key aspect of the work takes advantage of efficient pairwise comparisons between the training, test, and auxiliary instances to characterize their interaction within the vector space, and exploits it for improved classification in three ways. The first uses the local variability around the train and test instances to reduce false-alarms. The second assumes the instances lie on a low-dimensional manifold and uses the distances along the manifold. The third extracts relational features from a similarity graph where nodes correspond to the training, test and auxiliary instances. To quantify the merit of the proposed techniques, results of experiments in speaker verification are presented where only a single target recording is provided to train the classifier. Experiments are preformed on standard NIST corpora and methods are compared using standard evalutation metrics: detection error trade-off curves, minimum decision costs, and equal error rates.