A method for analyzing decision regions in Learning Vector Quantization algorithms

Abstract This paper presents a method for analyzing a vector quantizer. The proposed analysis is based on the statistics of codebook vector pairs in the quantizer and provides qualitative information of the classification borders and conditions for the use of the LVQ2.1 training algorithm. In addition, the connectivity of the decision regions of a nearest neighbor vector quantizer can be analysed. The basic concepts of the method are first visualized with an artificial two-dimensional example. Then, analyses of multiple 20-dimensional speech recognition codebooks are carried out.