Abstract: We propose a co-design approach for compute-in-memory inference for deep neural networks (DNN). We use multiplication-free function approximators based on l 1 norm along with a co-adapted ...