I. Introduction
Neural networks have been demonstrated to be successful in many territories, such as image processing and video processing, over the years [1]. Among these territories, some of the applications, such as self-driving, robot-assisted surgery, and medical diagnosis, can be mission-critical, and they have strict prediction accuracy requirements on the neural network models [2]. When the neural network models used in these applications are deployed on accelerators for the sake of higher performance and energy efficiency, the reliability of the underlying acceleration system becomes critical because hardware faults may lead to considerable wrong predictions [3] and system exceptions that can hardly be considered by the neural network model designers, which may cause catastrophic consequences.