A self-scaling procedure in unsupervised correlational neural networks

In neural networks, categorization is generally achieved by learning directly from the prototypes. However, in a natural setting the categories should emerge from learning from a set of exemplars instead of prototypes. Still even if the problem of learning from correlated items has been solved, the selection of the right size for a category remain an open question. In this study, we test the hypothesis that the introduction of a vigilance parameter which specifies the degree to which patterns must be similar in order to be considered exemplars of the same prototype can be implemented in a general correlational neural network. The results show that this is the case and the number of the resulting categories vary as a function of the value of the vigilance parameter It is thus concluded that such a vigilance parameter may constitute the key to self-scaling adaptation in unsupervised correlational neural network