Learning Sparse Combinatorial Representations viaTwo-stage Submodular Maximization