Deep Multi-View Learning via Task-Optimal CCA