A training method for enhancing neural network model generalisation

A training method for enhancing neural network model generalisation is proposed. In this method, a neural network is trained and tested alternatively on a training data set and a testing data set. Unlike in conventional neural network training where the training and testing data sets are fixed, the training and testing data sets swap roles continuously during network training. Training is terminated when the network prediction errors on both data sets cannot be further reduced. Application examples demonstrate that this neural network training strategy can significantly improve the neural tu network model prediction accuracy, especially the long range prediction accuracy.