Joint dynamic sparse learning and its application to multi-view face recognition

We propose a novel joint dynamic sparsity regularization for joint learning of multiple tasks (i.e., multiple observations of the same physical event by a set of homogeneous or heterogeneous sensors). The proposed method not only combines the strength of different tasks but also has the flexibility of selecting a set of different atoms for each task, with a class-wise constraint, which is more flexible and even crucial in many real-world scenarios. We develop an efficient learning algorithm for the joint dynamic sparsity using the accelerated proximal gradient descent. The proposed method is applied to a multi-view face recognition task and the experimental results on the public CMU Multi-PIE dataset verify its effectiveness.