Mixed-precision architecture based on computational memory for training deep neural networks | IEEE Conference Publication | IEEE Xplore