On hierarchical self-organizing networks visualizing data classification processes

This paper proposes a self-organizing neural network with hierarchical structure. In the forward phase of learning, the training data is propagated from the top-level neuron to one of the bottom-level neurons, and a combination of a parent neuron and its children, which the training data reaches, is a target for updating their weights. In the backward phase, weights of at least two neurons in such a combination are averaged, and weights of the parent are changed for the averaged weights. The proposed network adequately realizes polysemous data clustering, which yields multiple results, while sustaining the capability of data visualization.