PROPERTIES OF LEARNING IN ARTMAP

In this paper we consider the ARTMAP architecture for situations requi ring learning of many-to-one maps. It is shown that if ARTMAP is repea tedly presented with a list of input/output pairs, it establishes the required mapping in at most M(a) - 1 list presentations, where M(a) co rresponds to the total number of ones in each one of the input pattern s. Other useful properties, associated with the learning of the mappin g represented by an arbitrary list of input/output pairs, are also exa mined, These properties reveal some of the characteristics of learning in ARTMAP when it is used as a tool in establishing an arbitrary mapp ing from a binary input space to a binary output space. The results pr esented in this paper are valid for the fast learning case, and for sm all beta(a) values, where beta(a) is a parameter associated with the a daptation of bottom-up weights in one of the ART1 modules of ARTMAP.