Hardware Circuit Design And Simulation Of BP Neural Network For Cyclic Learning

With the advent of the information age,the requirements for data processing are getting higher and higher.The processor of the traditional digital computer performs information processing in a serial manner.In many cases,it cannot meet the needs of the present,and the artificial neural network is for Simulating a mathematical model of the human cranial nervous system,this parallel processing of artificial neural networks has led to hopes for the processing of massive amounts of data in the future.BP neural network is an artificial neural network trained by back propagation learning algorithm.At present,the software implementation of BP neural network is very mature and widely used.However,there are few researches on hardware implementation,so nowadays it is relatively In terms of software implementation,BP neural network hardware implementation can better reflect its parallel processing mechanism,which has very important practical significance.Firstly,by studying the basic principles of BP neural network,the BP neural network mathematical model is established.The main circuits needed to realize the whole BP neural network are determined from global to local,including weighted summation circuit,I/V conversion circuit and sigmoid activation function circuit.And learning circuit;then,respectively,the weighted summation circuit,the I/V conversion circuit and the sigmoid activation function circuit are simulated and tested to determine the implementation conditions and parameters of each circuit;secondly,the research status of the BP neural network hardware is realized in combination with the domestic and foreign,and the BP neural network is implemented.The application scope of the network has designed a loopable learning circuit,and the simulation proves that this circuit can continuously learn and refresh the weight through pulse control.Finally,from the local to the global,all the main circuits are combined to realize the whole BP neural network hardware.The circuit tests the effectiveness of the loopable learning circuit designed in this paper.The simulation results show that for the input of a single sample,the error of the simulation using the loopable learning circuit is only a few millivolts,while the relative error is only about 0.013%,and the relative error can be up to 4.78% without using the loopable learning circuit.For the input of multiple samples,the cyclic learning circuit is applied in the BP neural network circuit,and all the sample loop inputs can be synchronously controlled by the pulse during the control learning,and then the function fitting is taken as an example to obtain the relative error of the simulation.About 11.95%,which is similar to the relative error of 9.00% realized by C++ programming.It shows that BP neural network can be successfully implemented by hardware alternative software programming,and the validity of the loopable learning circuit is also verified.