基于dropout算法的卷积神经网络单粒子翻转容错方法

Single-event upsets fault tolerance of convolutional neural networks based on dropout algorithm

空间辐照干扰尤其是单粒子翻转(SEU)效应对神经网络芯片的正常稳定运行造成很大影响,它会导致存储在芯片SRAM存储器的权重参数随机发生比特位翻转,进而神经元的权重参数值发生变化,直接影响神经网络芯片输出的准确度。在分析现有的一些抗辐照干扰方法基础上,针对芯片硬件开销、恢复时间与处理速度的问题,利用软件仿真研究在不同比例权重参数出错的情况下神经网络的测试准确度,就结果准确度下降的情况,采用dropout算法构建新的网络框架,以一定概率屏蔽受到SEU影响的神经元。仿真实验结果表明,该方法可以提升受SEU干扰神经网络的准确度。

Space radiation interference especially the single event upset(SEU)effect can make a great impact on the normal and stable operation of the neural network chips,which can lead to the random bit flips of the weight parameters stored in the SRAM,and thus change the values of the neural weight parameters and the accuracy of the neural network chip outputs.In this paper,we analyze the commonly used radiation resistant methods,and try to overcome the problem of hardware consumption,recovery time,and processing speed.After the software simulation of the neural network accuracy with the different ratios of weight parameter errors,we adopt the dropout algorithm to construct the novel network framework to avoid the decline of accuracy.The algorithm can mask the neurons affected by SEU with some probability.The simulation experiment results show that this method can promote the accuracy of the neural network affected by SEU.