A Research Of Approximate Computing Technology For Low Power And Low Cost Neural Network Processor

How to reduce the power and cost of the neural network processor for IoT devices(such as wearable health monitoring devices,micro-intelligent sensing devices,etc.)is a key issue in the realization of artificial intelligence IoT(AIoT)technology.Adopting the approximate computing technologies,including binary neural network and emerging memory(such as memristor RRAM,near-threshold SRAM.),can greatly reduce the power and cost of the neural network processor.However,there is a large deviation in circuit parameters for RRAM or near-threshold SRAM,which leads to a large decrease in the accuracy of the neural network.How to solve the above problems of approximate computing technologies,while reducing the power and cost of the neural network processor and ensuring high accuracy,is the main research goal of this work.In response to the above issues,this paper has carried out the following research works:(1)By simulating the circuit parameters deviation of RRAM or near-threshold SRAM,a large number of experiments were performed on the weight matrices and accuracy of the binary neural network.It was found that the neural network after binary quantization has a certain fault tolerance,and different weight matrices have different effects on the accuracy of the neural network.(2)By analyzing the advantages and disadvantages of the emerging memory technology(including RRAM and near-threshold SRAM)and traditional SRAM memory technology,using the characteristics of binary neural network with certain fault tolerance,a mixed weight storage scheme is proposed,which maps the non-critical weight matrices to RRAM or near-threshold SRAM,and maps critical weight matrices to traditional SRAM,to greatly reduces hardware power and cost while ensuring sufficient accuracy.(3)For the search of non-critical weight matrices in binary neural network,two search algorithms are proposed.One is a weight search algorithm based on independent analysis,which searches all the weight matrices in the neural network one by one,to find the non-critical weight matrices and realize the mixed weight storage scheme.However,this method does not consider the correlation between the weight matrices,which may lead to a poor result.In response to this problem,a second search algorithm is proposed,that is,a weight search algorithm based on coupling analysis.This algorithm combines the binary particle swarm optimization algorithm to perform a combined search on noncritical weight matrices.Based on this,the algorithm initialization method is improved to speed up the search speed.This method can improve the search accuracy and further reduce the hardware power and cost.The above works have been verified by experiments,taking three neural networks of typical applications as examples.Through the weight search algorithm based on coupling analysis,under the condition that the accuracy loss of neural network is 1% and the error probability is 5%,non-critical weight parameters of more than 47%,around 86% and more than 95% are extracted from the three neural networks,respectively.With the mixed weight storage scheme,power savings of more than 35%,69%,and 77%,and hardware cost savings of more than 40%,78%,and 87% are achieved.