Learning and optimisation of hierarchical clusterings with ART-based modular networks

This paper introduces two optimization methods into learning of hierarchical clusterings with modular adaptive resonance theory (ART) networks. The aims are to reduce the complexity of trained networks and "clean up" the category prototypes during the learning process while maintaining the useful properties of hierarchical ART networks like fast and stable learning, and the ability to build category hierarchies incrementally. The experimental results demonstrate a significant reduction in category complexity as well as some improvement on a range of other metrics at a cost of varying amounts of additional training time. We suggest that scheduling the optimisation steps may be crucial in achieving an optimal trade-off.