Norimichi TSUMURA, Kazuyoshi ITOH and Yoshiki ICHIOKA

Osaka University, Department of Applied Physics, Yamadaoka 2-1, Suita, 565 Japan

(Accepted July 18, 1994)

A new artificial neural network that can reject strange patterns is presented. The new network is an improved version of the RCE network. Each cell in the last layer of the present network has two thresholds. The new threshold produces the smallest boundary that encloses all examples associated with the cell in the pattern vector space. The present network can reject strange patterns by using this boundary. The rejection of strange patterns in the present and RCE networks was investigated by practical experiments using digits sampled from X-ray films of the human chest.

Key words : strange pattern, rejection, RCE network, digits, classification