Self-organized significance analysis on automatically generated training data for neural networks

In many applications of neural networks, e.g. time series prediction or pattern analysis, training data are generated automatically out of large data sets. The problem is to determine the varying significance of the resulting training vectors concerning the given task in order to make appropriate decisions for the training phase. In this paper we propose a self-organized significance analysis based on a rareness assessment for each vector in the generated training data set. The resulting significance measure can be used to achieve considerably improved classification results for a wide variety of applications by systematically controlling training parameters like learning rate or frequency of presentation for each single vector.