Incremental On-line Clustering with a Topology-Learning Hierarchical ART Neural Network Using Hyperspherical Categories

Incremental on-line learning is an important branch of machine learning. One class of approaches particularly well-suited to such tasks are Adaptive Resonance Theory (ART) neural networks. This paper presents a novel ART network combining the ability of noise-insensitive, incremental clustering and topology learning at different levels of detail from TopoART with Euclidean similarity measures and hyperspherical categories from Hypersphere ART. As a result of the modified internal representations, several limitations of the original TopoART network are lifted. In particular, the presented network can process arbitrarily scaled values even if their range is not entirely known in advance.