Relation Organization of Initial SOM by Node Exchange Using Connection Weights

Self Organizing Map (SOM) is a kind of neural networks, that learns the feature of input data thorough unsupervised and competitive neighborhood learning. In SOM learning algorithm, every connection weight in SOM feature map are initialized at random to covers whole space of input data, however, this is also set nodes at random point of feature map independently with data space. Learning speed or learning convergence becomes slow is expected by this relation missing. As precedence research, I proposed the method that, initial node exchange by using a part of learning data, to improve the problem. Through this research, I thought the idea of initial node exchange must be effective even if learning data are not used. In this paper, here I propose new method, initial node exchange by using initial values of connection weights. This method is handled without the input from the outside. As a result of experiments, comparing with former method, new method is effective by about 5% smaller number of input data, but peek performance is about 6% inferior.