Connectionist networks qua graphs (Q1103413)

There is a lot of excitement in the field of artificial intelligence (AI) at the moment centering around the ideas of ``connectionism''. Connection networks are used to represent knowledge in terms of ``subsymbolic'' nodes (i.e. a single node does not by itself represent a conceptual entity, such as a dog, apple, Mary; any node may participate in a number of patterns of activation and each such pattern is a representation of a conceptual entity). The edges of these networks are arcs, with an associated numerical weight, whose role is to transfer ``activity'' from one node to another according to some function of the arc weights. The phenomenon of learning is typically modeled by adjusting arc weights according to some function of the network's desired and observed performance. The connectionist paradigm is seen as a promizing new approach to the realization of intelligent systems, and one that may be particularly amenable to formal analysis. This paper introduces connectionism, points out some of the major problems and argues that a graph theoretical approach to some of the recognized problems may prove fruitful.