Damage Analysis And Control Of Neural Network Accelerator Under Space Irradiation

Neural network technology is widely used in the fields of image recognition,data mining,computer vision,etc.More and more data to be processed is converted,and the network structure is more complex,which requires the consumption of a lot of computing resources.In order to ensure the effectiveness of the inference process of neural networks,Google,ARM and Cambrian have successively launched neural network accelerators.Such systems have very strict requirements on the working environment.The interference of the accelerator's storage unit,especially the singleparticle staggering effect,will cause the storage unit's parameters to be wrong.This error mapping into the neural network will cause the final output of the neural network to be biased.The characteristic analysis of the algorithm,combined with the fault tolerance of the neural network algorithm itself,expands the research to achieve the purpose of improving the fault tolerance of the neural network.The main research contents are as follows:1.Analyzing the dataflow acceleration method of the neural network accelerator and the main characteristics of the neural network inference process.The error of the weight parameter will affect the final output result as the data stream spreads.Combined with the effect of space irradiation on the single particle flip caused by the accelerator storage unit SRAM,a single event upset probability model was established,and the error injection of network parameters were annotated through software simulation.2.From the non-linear analysis of the activation function,the error concealment ability of different activation functions is analyzed,and the error concealment ability of the function that verifies the bilateral suppression effect through the analysis of error injection is stronger.It is further proposed to add a penalty term to the objective function to find the balance between the CNN optimal model and the fault-tolerant model Considering the characteristics of weight sharing of convolutional layers in CNN,the error of the parameters has a greater impact on the accuracy of the output results.This paper proposes to add a normalization layer after the convolution layer,and to verify its feasibility to improve the fault tolerance of the network through error injection experiments And through error analysis of RNN,it is verified that the normalization algorithm can still improve the fault tolerance of the network.