| Summary: | With the huge research effort into connectionist systems that has taken place over the last decade a debate has developed as to whether the more traditional Artificial Intelligence (AI) paradigm of symbolism or the connectionist paradigm offers the way ahead to developing high level cognitive systems. Central to the debate are issues of representation. Traditional AI has spent many years developing representation languages and representation has long been seen as essential for the development of intelligent systems. Early connectionists have tended to rely on the notion that a network of simple processing units will develop adequate internal representations as a by product of learning. Indeed, with connectionism it would appear on first sight that the development of a representational formalism is somewhat intractable when knowledge is implicit in a distributed pattern of activity. Contrary to this view, some connectionists have agreed with the traditionalists that the mechanism of represnetation must support compositional construction and be understood. Some connectionists would even go as far to say that the representation mechanism should be understood to the point whereby an explicit or easily read description of the knowledge held by a network can be given. This thesis presents some of the key issues which arise when attempting symbol style representations with connectionist architectures. A number of connectionist techniques are reviewed. The emphasis of this thesis is on the presentation of a model that provides a simplified version of a connectionist system that was developed to represent symbol structures. The model is the result of the research reported herin and provides an original contribution in a number of important areas. The model has the benefit of allowing very quick derivation of connectionist representatiions, unlike the slow training environments of a pure network implementation. The model provides a mathematical framework that gives insight into the convergence behaviour of the technique it proposes and this framework allows a statement to be made about generalisation characteristics. The model has immediate practical use in supplying connectionist representations with which to experiment and provides a conceptual cehicle that should assist with the development of future techniques that tackle representation issues.
|
|---|
