Learning And Interpreting Deep Representations From Multi-Modal Data