Evaluation and Mitigation of Single Event Upset on the Accuracy of Neural Network Accelerators

Deep neural networks (DNNs) are widely used in the field of satellite remote sensing, which needs to be deployed in neural network accelerators to meet the demand for high energy efficiency and low computing latency. However, the high-energy particles, such as the single event upset (SEU), directly lead to the accuracy reduction of DNNs, which has not been systematically evaluated and mitigated. In this paper, the evaluation and mitigation of SEU on the accuracy of neural network accelerators are proposed. The simulator of SEU on DNNs is proposed, and SRAM and registers are fault injected. The quantized neural network parameters are analyzed from the bit level and the layer level, and we mitigate it with the bit & level triple modular redundancy (TMR) technique. Experimental results show that the DNN accuracy recovers in VGG-16 and ResNet-18 with 12.51% area and 12.47% energy overhead.