Abstract
A major feature of the self-organizing map (SOM) is a topology-preserving projection developed through training, and it has been accomplished based on the convergent manner that makes the best-match neuron (BMN) respond stronger than before. In contrast, a new training method based on the divergent manner is proposed in this study. Its essence is as follows: i) the worst-match neuron (WMN) is chosen in the competitive layer, and ii) its reference vector is updated to leave from the input vector more than before. As a result of computer simulations, it is confirmed that the proposed method also develops similar feature maps successfully.
