Reliability Enhancement of Neural Networks via Neuron-Level Vulnerability Quantization

Neural networks are increasingly used in recognition, mining and autonomous driving. However, for safety-critical applications, such as autonomous driving, the reliability of NN is an important area that remains largely unexplored. Fortunately, NN itself has fault-tolerance capability, especially, different neurons have different fault-tolerance capability. Thus applying uniform error protection mechanism while ignore this important feature will lead to unnecessary energy and performance overheads. In this paper, we propose a neuron vulnerability factor (NVF) quantifying the neural network vulnerability to soft error, which could provide a good guidance for error-tolerant techniques in NN. Based on NVF, we propose a computation scheduling scheme to reduce the lifetime of neurons with high NVF. The experiment results show that our proposed scheme can improve the accuracy of the neural network by 12% on average, and greatly reduce the fault-tolerant overhead.