Efficient SOM learning by data order adjustment

Abstract

Kohonen's self organizing maps (SOM) is a kind of neural network that the algorithm learns the feature of input data thorough unsupervised and competitive neighborhood learning. SOM is mapped from a high dimensional space onto a two dimensional space, so it can visualize the high-dimensional information to the map. In the SOM's learning algorithm, there are many factors to aggravate the computational load and a competition to be declared the winner. We think it is a major factor at the beginning of learning process that SOM's map is changing dynamically and widely and the learning dynamics depends on the distance of each input data. Thus we suppose that, by adjusting the data order, the competition must be reduced and the learning convergence must become faster. In this paper, we discuss the efficient learning by data order adjustment, and compare it with the conventional method. We achieved a maximum 9% improvement.