Speaker: Ren-Cang Li (Hong Kong Baptist University and University of Texas at Arlington) Title: A Multiview Contrastive Trace Ratio Learning Date: 4pm Friday, August 12, 2022 Venue: https://us02web.zoom.us/j/84923348751?pwd=WkV0MUxkcS9KS0pOLytkUWFjaS9ZQT09 회의 ID: 849 2334 8751 암호: 104071 Abstract: One of the guiding principle in designing machine learning models is to highlight one characteristic of interest in data while diminish others, or equivalently to contrastively pitch one characteristic against others. Ratios and differences are two universal ways for doing that. This talk concerns a multiview contrastive learning model in the form of a trace ratio optimization problem over the product of Stiefel manifolds. The model has a built-in contrastive parameter that can be used to flexibly control the level of contrastiveness. For example, it includes Fisher’s linear discriminant analysis (LDA) and a recent orthogonal canonical correlation analysis (OCCA) as special cases. We will explain how to efficiently solve the model with alternating iterative schemes that repeatedly solve certain nonlinear eigenvalue problems with eigenvector dependency by a novel self-consistent-field (SCF) iteration, analyze the convergence of the method, and then demonstrate its instantiated concrete models on real world multiview data sets. Numerical results will be presented to show the efficiency of the proposed numerical methods and effectiveness of the new multi-view subspace learning model.