NN/II: improved version of a network for large-scale pattern recognition tasks
Hajime Kita, Hirokazu Masataki, Yoshikazu Nishikawa · 1991
A network architecture NN/I which divides and learns environments was previously proposed for pattern recognition. NN/I divides a task of learning input-output relations into several subtasks adaptively, and then learns each subtask in a part of the network, i.e., subnetwork, by means of the error backpropagation algorithm. The results of an application of NN/I to recognizing handwritten Japanese characters are summarized, and its faults in generalization ability revealed in the application are analyzed. NN/II, which accommodates various novel devices to remove the faults in NN/I, is introduced, and its satisfactory ability is demonstrated through an application to character recognition.>