A comparison of two ART-based neural networks for hierarchical clustering

The paper compares two modular neural network architectures, built up of adaptive resonance theory (ART) networks, that can develop stable two-level hierarchical clusterings of arbitrary sequences of binary input patterns. In particular, it contrasts the typical class hierarchies that the networks found on a machine learning benchmark database. It is proposed that the main difference between the two clusterings are the direct consequence of the existence or absence of an internal feedback mechanism and explicit associative links between a higher-level class and its sub-classes.