Self-organizing continuously-overlapping map is shown to have ability to detect the first and second nonlinear principal components. This is an extended version of the self-organizing overlapping mapping. The model was applied to FFT data of sound, and some others. These data are characterized by a combination of two kinds of features, such as the pitch and the quality of tone. The model has two self-organizing layers. One layer extracts and maps continuously one feature, and the other layer does the same with respect to the other feature. The ability of generalization depending on data structure is demonstrated. Comparison to Kohonen's SOM is also discussed.
Recently, modular networks have been used to try to solve efficiently multiclass classification problems. However, the rejection rate on patterns of unlearned classes is usually very low. Moreover, when new classes are later added, old modules in the usual modular network need to be re-trained. A modular network proposed in this paper has RBF output units and an algorithm for incremental learning that improve these points. The results of computer simulations showed that the model achieved higher rejection rates on patterns of unlearned classes than the usual modular networks.