Design & implementation of the parallel-distributed neural network ensemble

Neural network ensemble is a recently developed technology, which trains a few of neural networks and then combines their prediction results. It significantly improves the generalization ability of neural network system and relieves the trial-by-error process of tuning architectures. However, it is time-consuming. In order to overcome the disadvantage, a parallel-distributed neural network ensemble named PDNNE is proposed in this paper. The design and implementation of the PDNNE are presented through discussing the main issues such as partitioning, communication, and the component neural network. The experiments show both the generalization ability and time efficiency are significantly improved.