Unsupervised learning of neural networks for separation of unknown data

In neural networks, the learning scheme is very important and, basically, is divided into supervised learning and unsupervised learning. If one would like to classify a set of data, the statistics of which are not known, then one cannot apply an ordinary supervised learning scheme. On the other hand, if one can embed a relation between input data and teaching signals into an evaluation function, one can allow neural networks to learn the relation. In this paper, the authors propose an unsupervised learning scheme and an evaluation function that realizes a classification of unknown data. Some simulation results are also shown.