I. Introduction
Deep neural networks (DNNs) have been rapidly developing in various fields of machine-learning applications [1], [2]. The size of modern DNN models is growing in order to obtain higher recognition accuracy and robustness for a wide range of applications [3], [4]. These DNNs require a large memory footprint and they consume high energy. To implement an accelerator for this powerful computing on limited energy and resources, it is important to combine a new device technology and an optimization algorithm.