| Deep neural network(DNN)requirs a lot of operations and consumes a lot of energy and hardware resources.The memristor has the excellent characteristics of small area and low power consumption,and its resistance varies with the current flowing through it.Therefore,it is very suitable to realize the synapse of neural networks.Memristor neural network(m NN)uses memristor crossbar array(MCA)to store weights and perform point multiplication operations,which greatly improves the computing performance.However,due to the immature manufacturing process of the memristor,there are a large number of faults in MCA.m NN has serious problems in the recognition performance and service life.At present,though some hardware redundancy technologies have been proposed at a high design level,the fault tolerance of m NN needs to be further improved.This thesis proposes a weight-level redundancy(WLR)method for stuck-at faults(SAFs),which are the most common faults in MCA.Hardware redundancy structure at a low level is designed to obtain high fault tolerance.In this method,each weight is stored using memristors that add R time’s redundancy.When some of the memristors fail,other fault-free memristors can be used to make up for the deviation of the weight.Theoretical analysis shows that when the fault rate of the memristor is p,the failure rate of adding 1,2 and 3 times redundant weight units is reduced to about 3p/2,5p~2/2,35p~3/8 respectively.Neural networks have inherent fault tolerance and can tolerate a small number of faults.We also further improve the recovery rate of m NN recognition accuracy through retraining.Finally,two-layer and three-layer fully connected networks are used to verify the effectiveness of the proposed method WRL on MNIST data sets.Simulation results show that,compared with previous methods,WLR has a higher average accuracy recovery rate.Even if the fault rate of a memristor is very high,an average accuracy recovery rate of not less than 99.20%can be obtained by adding 1 time redundancy for20%fault rate of the two experimental models and 3 times redundant WLR at 40%fault rate.After retraining,the average accuracy recovery rate is further improved,and the corresponding recovery rate was increased to no less than 99.78%. |
