Initializing probabilistic linear discriminant analysis

Component Analysis (CA) consists of a set of statistical techniques that decompose data to appropriate latent components that are relevant to the task-at-hand (e.g., clustering, segmentation, classification, alignment). During the past few years, an explosion of research in probabilistic CA has been witnessed, with the introduction of several novel methods (e.g., Probabilistic Principal Component Analysis, Probabilistic Linear Discriminant Analysis (PLDA), Probabilistic Canonical Correlation Analysis). PLDA constitutes one of the most widely used supervised CA techniques which is utilized in order to extract suitable, distinct subspaces by exploiting the knowledge of data annotated in terms of different labels. Nevertheless, an inherent difficulty in PLDA variants is the proper initialization of the parameters in order to avoid ending up in poor local maxima. In this light, we propose a novel method to initialize the parameters in PLDA in a consistent and robust way. The performance of the algorithm is demonstrated via a set of experiments on the modified XM2VTS database, which is provided by the authors of the original PLDA model.