Research On Defect Tolerance Method For Memristor Neural Network Accelerator

Deep neural network has been widely used because of its excellent predictive performance.However,the traditional computing platform cannot efficiently implement the calculation of neural network due to the low parallelism and the problem storage wall.The core calculation of neural network is matrix vector multiplication.Memristor crossbar is regarded as a more promising solution for the design of neural network accelerator because it can store weight matrix and realize matrix vector multiplication with constant computational complexity.However,there is no way to avoid hardware defects in the manufacture of memristors.The electrical manifestation of a defective memristor is fault.The faulty memristor can not correctly represent the weights of the trained network,which will reduce the classification accuracy of the neural network deployed on the memristor crossbar.The method of weight mapping is a low cost method of defect tolerance.By adjusting the mapping of weights to memristors,we can improve the classification accuracy of the neural network deployed on the memristor crossbar.Most of the existing fault-aware methods of weight mapping are based on the row granularity,which limits the search space for the optimal solution.In order to solve the above problems,we propose two weight mapping methods at operation unit level to achieve a finer-grained optimization.We divide the memristor crossbar into many small operating units.These operations units can be activated independently and perform calculations.The first mapping algorithm we proposed improves the network accuracy by adjusting the mapping inside each OU.Firstly,we achieve the preliminary mapping of the weight matrix to the memristor crossbar in weight row granularity.We then further fine-tuned the mapping of weights within each OU to achieve finer grained optimization and better tolerance for defects and failures.On the other hand,after the weight mapping in the OU is adjusted,the position of the elements in the input vector needs to be adjusted as the weight position in the row dimension is adjusted.Therefore,different OUs require different input vectors for their calculations.The increase in the number of input vectors not only requires more storage space,leads to more complex control circuits,but also reduces the computational performance of the crossbar.We propose a unified input algorithm to reduce the number of input vectors.The algorithm quantifies the similarity of input vectors of different OUs according to Euclidean distance,and unifies the closest input vectors into a single vector and is used in the calculation of these OUs.Reducing the number of input vectors at the expense of an acceptable loss of accuracy.The second mapping algorithm is the OU granularity weight mapping algorithm based on the KM improved algorithm,which takes into account the influence of gradient and deviation of weight.The mapping algorithm optimizes the sub-weight matrices corresponding to the OUs in the same column uniformly,so that each column OU only needs to store one input vector,which reduces the storage of input vector.In this paper,three kinds of deep neural networks are mapped to a memristor crossbar by using the two proposed algorithms.The experimental results show that,when the defect rate is 20%,the accuracy of Alex Net,VGG16 and VGG19 networks mapped using the first mapping algorithm improve 11.07%,3.91% and 2.24% compared with the row granularity mapping method.At the same time,the number of input vectors reduces10% by the unified input algorithm compared with the situation without considering the number of input vectors optimization,and the loss of network precision is less than 1%.The experimental results show that the accuracy of the network mapped by the second mapping algorithm is the highest.When the defect rate is 20%,the accuracy of the Alex Net,VGG16 and VGG19 networks mapped by this algorithm improve 1.2%,1.31%and 0.8% compared with that network mapped by the first mapping algorithm.The number of input vectors is also less than the number of input vectors optimized by the first algorithm.