Research on Mixed-Precision Quantization and Fault-Tolerant of Deep Neural Networks

As deep neural networks become more and more common in mission-critical applications, such as smart medical care, drones, and autonomous driving, ensuring their reliable operation becomes critical. The data in the hardware memory is susceptible to bit-flip due to external factors, which leads to a decrease in the inference accuracy of the deep neural network deployed on the hardware. We solve this problem from the perspective of the deep neural network itself, We use a reinforcement learning algorithm to search for the optimal bit width for the weights of each layer of the deep neural network. According to this bit width strategy, the deep neural network is quantified, which maximizes the limitation of data fluctuations caused by bit-flip and improves the fault-tolerance of the neural network. The fault-tolerance of the network model compared with the original model, the solution proposed in this paper improves the fault-tolerance of LeNet5 model by 8.5x , the fault tolerance of MobileNetV2 model by 15.6x , the fault tolerance of VGG16 model by 14.5x , and the accuracy decreases negligibly.