Evaluating the Impact of Fault-Tolerance Capability of Deep Neural Networks Caused by Faults
Yung-Yu Tsai, Jin-Fu Li · 2021
Deep neural network (DNN) is one effective technique used for the artificial intelligence applications. A DNN consists of a large amount of neurons arranged in a form of multilayers. Typically, a DNN has over-provisioning neurons such that it has inherent fault-tolerance capability. However, how to evaluate the fault-tolerance capability of DNNs is a question. In this paper, we propose a simulator to estimate the loss of inference accuracy due to the faults in a DNN model or the memory faults in hardware accelerator. The simulator is implemented based on the platforms of Keras and Tensorflow. It can evaluate the fault-tolerance capability of a DNN at model and hardware levels.