Author: Guszti Bartfai
Source: GZipped PostScript (57kb); Adobe PDF
(267kb)
This paper presents a generic modular neural network architecture built up of Adaptive Resonance Theory (ART) networks. The network has a more powerful knowledge representation than its component ART networks in that it is capable of representing {\em class hierarchies} as opposed to simply partitioning the input data set. At the same time, the network retains the main properties of ART networks like fast and stable incremental learning, parallel search and so on. Three different implementations of the generic network (SMART, HART-J and HART-S) are described, each being capable of unsupervised learning of hierarchical clusterings of arbitrary sequences of input patterns. Each hierarchical ART network produces slightly different clusterings due to the differences in the interactions between the component ART modules and the input each module receives. Two examples, produced by computer simulations, illustrate the developed 2 and 3-level clusterings on a machine learning benchmark dataset. The architectures are also general enough that the types of ART modules can be changed (e.g. from binary ART to Fuzzy ART) --- without affecting the main properties of the network --- to suit the problem domain better. We also ask questions about how this model might be extended and/or modified to make it suitable for supervised learning, to make better use of the structured knowledge present in the network, to make the resulting clusterings more robust, and to create new ART modules during learning. We speculate that all these could be achieved by implementing appropriate control mechanism in the networks.