Reliable classification by double hyperspheres in pattern vector space

摘要

A new hyperspherical classifier that performs reliable classification is presented. The new classifier was inspired by the RCE network. The present classifier has two important features; (1) cells in the last layer of the two-layered architecture have two thresholds, and (2) weight vectors are modified according to training patterns. The two thresholds produce double hyperspheres in the pattern vector space. The double hyperspheres determine regions of rejection. Patterns in these regions are rejected and not classified in any classes. The weight vectors are optimized by modification. Results of practical experiments are presented to compare the reliability of the new classifier with that of the RCE networks.