Centroid-driven grouping based ART network for fast, incremental, and multiple clustering

Abstract

Adaptive Resonance Theory (ART) is an unsupervised clustering algorithm that enables fast and incremental learning. However, there are three limitations of fusion ART which is most frequently used ART network. 1) Basically, fusion ART divides the data into hyper-rectangles called category boxes. If the input data belongs to overlapping area of multiple category boxes, it can not be clustered properly. 2) Because fusion ART can only divide data into hyper-rectangles, if the data consists of arbitrarily shaped clusters, it can not be clustered properly. 3) Finally, multiple clustering, many-to-many mapping is not possible. In this paper, we analyzed the limitations and proposed multi-channel Centroid-driven Grouping based ART (CG-ART) to solve the limitations. In the multi-channel CG-ART, the concept of centroid is introduced and the grouping process is added. Each cluster has a centroid vector, and a cluster with the centroid vector closest to the input vector is selected. This allows proper clustering even when the input vectors belong to overlapping area of multiple category boxes. In the grouping process, the degree of overlapping of the category boxes is measured. When the overlapping degree is larger than the user-defined grouping parameter, clusters are grouped together. This allows clustering data consisting of arbitrarily shaped clusters. In addition, multiple clustering is also possible because clustering results are different for each user-defined grouping parameter. In the latter part of this paper, CG-ART is applied to episodic memory. Episodic memory is an architecture that can learn, recall, and retrieve episodes that are the sequence of events. In this paper we proposed Deep CG-ART that can recall multiple episodes by applying the proposed algorithm to Deep ART, which is a representative episodic memory architecture. In the simulation chapter, we measured and compared the performance of CG-ART in 11 benchmark data sets. Performance was measured by normalized mutual information and clustering purity. In the application chapter, Deep CG-ART was implemented to learn 38 episodes and we confirmed that it recalls multiple episodes.