Conference paperMixed-precision architecture based on computational memory for training deep neural networks