PURPOSE: To converge learning within a short processing time by setting up the contents of a weight vector so as to be a representative vector of respective cluster obtained at the time of clustering a learning vector belonging to a category and executing a teacher existence learning.
CONSTITUTION: Self-organization learning is executed in an optional category at first, a learning vector group belonging to the category is clustered and a weight vector relating to the category is initialized so as to become the representative vector of the obtained clusters. The initialization of weight vectors relating to all the categories including respective learning vectors constituting the learning data group has been completed, teacher existence learning is executed by using the initial values of the weight vectors, and when the value of similarity between the learning vector calculated at the time of teacher existence learning and the weight vector is included within a prescribed range and the learning is converted, the learning of the neural network is completed.
UEDA TORU
ARAMAKI TAKASHI
ISHIZUKA YASUSHI
