Multiple kernel active learning for facial expression analysis

Abstract

Multiple Kernel Learning (MKL) approaches aim at determine the optimal combination of similarity matrices (since each representation leads to a different similarity measure between images, thus, kernel functions) and the optimal classifier simultaneously. However, the combination of "passive" kernels learning scheme limits MKL's efficiency because side information is provided beforehand. A framework of Multiple Kernel Active Learning (MKAL) is presented in this paper, in which the most informative exemplars are efficiently selected by min - max algorithm, the margin ratio is used for querying next instance. We demonstrate our algorithm on facial expression categorization tasks, showing that the proposed method is accurate and more efficient than current approaches. © 2011 Springer-Verlag.