Abstract
This paper proposes a new class generation method for pattern classification problems. The approach enables estimation of a posteriori probability for learned and unlearned classes using a neural network, and allows evaluation of unlearned data distribution structures with a self-organization map (SOM) based on the probabilities estimated. By reconstructing the network structure dynamically, new classes can be generated from data of undefined classes via network training using given learning and evaluation samples. The method can be applied to various pattern discrimination problems such as electromyogram (EMG) classification. In the experiments reported here, classifiability with the proposed method was demonstrated using EMG patterns. The results showed that the technique provides a high level of performance for learned, unlearned and new class discrimination.
