Adaptive Mapping Networks for Visualizing Topological Structure of Data on Category Maps

Abstract

This paper present Adaptive Mapping Networks (AMNs) as an adaptive and incremental method to learn series data for visualizing on a category map. The architecture of AMNs comprises three modules: codebook module, labeling module, and mapping module. The codebook module quantizes input data as codebooks of low-dimensional feature vectors using Self-Organizing Maps. The labeling module creates labels as a candidate of categories based on the incremental and adaptive learning of Adaptive Resonance Theory. The mapping modle visualizes spatial relations of categories on a category map using Counter Propagation Networks. AMNs actualize supervised learning and unsupervised learning to change its network structures. The experimental results using open datasets of two types show the recognition accuracy of our method is superior than that of the existing method. Moreover we present the usefulness of visualizing functions using category maps.