Is optimization the right language to understand deep learning?