Transition Learning by Negative Correlation Learning

Abstract

The idea of transition learning is to apply negative correlation learning with one error learning function for a certain time, and then to switch to another learning error function. Because of the different learning functions between the two periods, the learning hehaviors are expected to have a sudden change in transition learning. On one hand, negative correlation learning with the lower λ might learn too well the training data while generating less negatively correlated neural networks. On the other hand, negative correlation learning with the larger λ might not be able to learn well the training data, but be capable of generating highly negatively correlated neural networks. With transition learning, the ensembles could have both the good performance and the diverse individual neural networks.