基于Spark的改进并行BP算法

BP (Back Propagation) neural network is a multilayer feedforward neural network with error back propagation, and it is one of the most popular neural network models. The traditional BP algorithm has a slow convergence rate, long training time, often caught by a local minimum and easy to overfit. In this paper, we improved the traditional BP algorithm in many aspects,such as the self-adaptive learning rate, the momentum factor, the mini-batch gradient descent method, the cross entropy cost function, the early stop criteria and so on, and then implemented the improved algorithm based on spark platform. Spark is a memory-based parallel computing framework, and it is good at dealing with iterative computation, so it is very suitable for the parallelization of BP algorithm. Compared with the experimental results of multi-layer perceptron classifier of Mllib, we found that the BP algorithm in this paper can ensure the accuracy of the classification, and the convergence speed is increased by at least 30%.