Multiview Correlation Feature Learning with Multiple Kernels

Recent researches have shown the necessity to consider multiple kernels rather than a single fixed kernel in real-world applications. The learning performance can be significantly improved if multiple kernel functions or kernel matrices are considered. Motivated by the recent progress, in this paper we present a multiple kernel multiview correlation feature learning method for multiview dimensionality reduction. In our proposed method, the input data of each view are mapped into multiple higher dimensional feature spaces by implicitly nonlinear mappings. Three experiments on face and handwritten digit recognition have demonstrated the effectiveness of the proposed method.