Research On Parallelization Of BP Neural Network Based On MapReduce

BP neural network is one of the most widely used artificial neural networks at present.In order to improve the application of BP neural network in big data,it is necessary to design parallel neural network.The parallel research of BP neural network based on MapReduce can process large-scale data and improve training speed,but there are still problems such as slow convergence of BP neural network and easy to fall into local convergence,etc.,and the local convergence weight matrix generated by Map phase training in Reduce The global convergence weight matrix obtained after the processing cannot guarantee that MapReduce training will proceed in the direction of global convergence.This paper proposes two different solutions to the above problems.One is to optimize the initial weights and thresholds of the BP neural network in the Map phase,that is,the weights and thresholds optimized by the particle swarm optimization algorithm can improve the convergence speed of the BP neural network and thus reduce the training time of the model.One is to use genetic algorithm instead of average value in reduce phase,carry out global optimal search for the weight matrix collected in Map phase,and substitute the optimal matrix in the next training,which can not only reduce the number of iterations,but also make the training in the direction of global convergence.In this paper,two improved models and traditional models were trained respectively.The experimental results show that both methods can reduce the training time and improve the training accuracy.