Research On The Realization Of BP Neural Network Based On FPGA

As one of the important research directions in the field of artificial intelligence,neural network is widely used in pattern recognition,image processing,robot control and other fields.The research of neural network includes network structure,network training and network application.Most of the network training is implemented by software,but the slow speed and poor parallelism restrict its application.Therefore,the technology of using hardware to realize neural network has emerged,among which using FPGA to realize neural network is one of the research directions.FPGA has the advantages of high parallel processing speed,flexibility and reliability,short design cycle,low cost and so on.The operation processing of neural network also has parallel characteristics.Therefore,the realization of neural network with FPGA can give full play to its advantages,greatly improve the training speed and then expand the application field of neural network.In this paper,the method of training and recognizing BP neural network with FPGA is studied.Taking the handwritten digit recognition system as the research object,a threelayer BP neural network is realized with FPGA,which can be used to recognize handwritten digits.Handwritten digital samples come from MNIST database,which is used as a prototype to build a three-layer BP neural network structure.According to the training and recognition algorithm of BP neural network,the formula and algorithm of BP neural network based on FPGA are deduced,including forward operation,backward operation,training cycle,etc.The whole code is designed with Verilog language,the logic synthesis is carried out with Quartus II 13.0,and the function simulation is carried out based on Modelsim-Altera.The results are correct and reliable.In this paper,the principle of BP neural network is analyzed.Formula deduction and implementation algorithm,system structure and module implementation,test scheme,simulation results and analysis,data fixed-point processing,activation function and derivative implementation method,matrix operation principle are given.In this paper,the training and recognition of BP neural network based on handwritten digit recognition is realized by using FPGA.The results can be transplanted to other BP neural network training and recognition,and lay a research foundation for further research on the implementation of other types of BP neural network in the field of FPGA.Further improvement can be applied to real-time,small embedded AI system.