moDNN: Memory optimal DNN training on GPUs | IEEE Conference Publication | IEEE Xplore