Comparative study on mirror image learning (MIL) and GLVQ

Abstract

The effectiveness of a corrective learning algorithm MIL (mirror image learning) is comparatively studied with that of GLVQ (generalized learning vector quantization). Both MIL and GLVQ were proposed to improve the learning effectiveness beyond the limitation due to independent estimation of class conditional distributions. While the GLVQ modifies the representative vectors of a pair of confusing classes when recognizing each learning pattern, the MIL generates a mirror image of a pattern which belongs to one of a pair of confusing classes and increases the size of the learning sample to update the distribution parameters. The performance of two algorithms is evaluated on handwritten numeral recognition test for IPTP CD-ROMI. Experimental results show that the recognition rate of the projection distance classifier is improved from 99.31% to 99.40% by GLVQ and to 99.50% by MIL, respectively.