Undesirable results of SOM learning and the feature of learning data

Kohonen´s self organizing map (SOM) involves neural networks, for which an algorithm learns the feature of input data through unsupervised, competitive neighborhood learning. In many cases of SOM learning, if the data make classes in input data space with similar density, similar shape, and similar size, corresponding classes in feature map also formed to similar shape and similar size. In the experiments, however, we found undesirable learning results, that corresponding classes in feature map formed to different shape and different size one another. In this paper, we investigate what kind of learning data set, which feature of learning data causes undesirable results.