Low-Rank Bilinear Classification: Efficient Convex Optimization and Extensions

Abstract

In pattern classification, it is needed to efficiently treat not only feature vectors but also feature matrices defined as two-way data, while preserving the two-way structure such as spatio-temporal relationships. The classifier for the feature matrix is generally formulated in a bilinear form composed of row and column weights which jointly result in a matrix weight. The rank of the matrix should be low from the viewpoint of generalization performance and computational cost. For that purpose, we propose a low-rank bilinear classifier based on the efficient convex optimization. In the proposed method, the classifier is optimized by minimizing the trace norm of the classifier (matrix) to reduce the rank without any hard constraint on it. We formulate the optimization problem in a tractable convex form and provide the procedure to solve it efficiently with the global optimum. In addition, we propose two novel extensions of the bilinear classifier in terms of multiple kernel learning and cross-modal learning. Through kernelizing the bilinear method, we naturally induce a novel multiple kernel learning. The method integrates both the inter kernels between heterogeneous reproducing kernel Hilbert spaces (RKHSs) and the ordinary kernels within respective RKHSs into a new discriminative kernel in a unified manner using the bilinear model. Besides, for cross-modal learning, we consider to map into the common space the multi-modal features which are subsequently classified in that space. We show that the projection and the classification are jointly represented by the bilinear model, and then propose the method to optimize both of them simultaneously in the bilinear framework. In the experiments on various visual classification tasks, the proposed methods exhibit favorable performances compared to the other methods.