An adaptive vigilance parameter tuning for ARTMAP

Abstract

Adaptive Resonance Theory (ART) is designed to achieve continual learning. ART trains clusters based on the similarity between input data and clusters. ARTMAP is a supervised learning method based on ART, and has excellent continual learning capability to adapt to new data while maintaining past learning content. On the other hand, its drawback is that the vigilance parameters given by the user in advance have a significant impact on the classification performance. In this study, we propose an approach to adaptively compute a vigilance parameter based on the diversity of training data distributions that are input sequentially. Numerical experiments on real-world data show that while the conventional ARTMAP requires optimal parameter settings for each dataset, the proposed method achieves comparable or better continual learning performance without parameter settings.