Kohonen's Self Organizing Map (SOM) is a kind of neural networks, that learns the feature of input data thorough unsupervised and competitive neighbourhood learning. According to SOM learning algorithm, SOM learning speed is affected by learning data and initial value of feature map. By improving conventional initialization method of feature map, the influence to the learning speed is expected. In conventional initialization process, very connection weights in SOM feature map are initialized to random values, and this is also set nodes to random point of SOM feature map independently with data space. Because some linkage between them is created when learning is completed, It is expected that no linkage at initialization must be influenced to the learning speed. In this paper, here I proposed a new method, node exchange of initial SOM feature map using learning data, and a new measure of convergence, the average of the move distance of all output nodes. As a result of experiments, the effect is sufficient by performing exchange process by about 5% of the whole learning data, and the average of the move distance of all nodes was drastically shortened about 64% through about 45%, and learning speed is improved that it becomes about 55% by this method.
View full abstract