A feature-weight detector neural network and its application

A feedback neural network model with memory connections for classification and weight connections for selection is proposed. After training, a memory vector is interpreted as a prototype of a feature pattern, and a weight vector represents importance of feature variables to the corresponding feature pattern. The proposed neural network has a simple network architecture and high learning speed. Moreover, the obtained knowledge can be described by natural language. The technique is applied to the IRIS data: the two effective feature variables were extracted, and the corresponding number of errors, is almost the same as using four feature variables.