End-to-end Learning of LDA by Mirror-Descent Back Propagation over a Deep Architecture