Adaptive resonance theory and the classical leader algorithm

Summary form only given. The classical leader algorithm described by J.A. Hartigan was compared to binary adaptive resonance theory (ART1). The aim is to make the functionality of ART more transparent and in the process shed light on some of its benefits and limitations. ART is characterized by unsupervised learning and identification of novel patterns as distinguished from refinements to existing pattern cluster prototypes. Lippmann has already noted the close similarity between these two clustering methods, and this relationship was verified in the work presented by implementing the leader algorithm with two datasets: seventeen alphabet characters and ten geometric shapes. The distance and threshold parameters of the leader algorithm were replaced with the template matching and vigilance parameters of ART. The focus is on the ramifications of the demonstrated similarity, especially input order dependency. It is recognized that the leader algorithm is remote programming, whereas ART is a biology-based `learning machine´