Relative effectiveness of training set patterns for backpropagation

Backpropagation (BP) is a gradient-descent method to search for optimal parameter settings in a neural network model. This learning process can be split into three stages according to the behavior of the errors produced. The analysis of the different behaviors in these stages shows the existence of poorly trained patterns which have great influence on the performance of the BP mode. The benefit of considering the relative effectiveness of training patterns is investigated using two modified BP training procedures. The versions are based on information about relative importance. The results show that the learning speed and generalization ability are improved