An auto-augmenting neural network architecture for diagnostic reasoning

Abstract

During the past two decades, there have been tremendous efforts in the field of artificial intelligence(AI) to develop diagnostic reasoning models, including statistical pattern classifiers based on Bayes' theorem, yule-based systems employing deductive reasoning, and associative network systems which are built on inductive reasoning paradigms. Although each of those models has its strengths within relatively well suited application domains, none provide adequate mechanisms to support the interactions among knowledge structures which are necessary to capture the generative capacities of human diagnosticians in novel situations. Achieving such interactions has been one of the greatest difficulties associated with implementing models of diagnostic decision-making that can reason in the presence of imprecise or incomplete information, using traditional AI representations like frames [Minsky, 1975] and production rules [Newell & Simon, 1972]. Throughout this period, however, there has been an unabated interest in developing radically different models of human memory and cognition. , This impetus for a new perspective to cognition can be largely attributed to a desire to overcome the shortcomings of existing symbol processing models to emulate human decision-making in terms of its underlying structure, representation and processing of knowledge, but more importantly, to better understand its evolution with experience over time. Not surprisingly, there has been a resurrection of a previously abandoned approach to studying the mind, called "Parallel Distributed Processing", "Connectionism", or "Neural Networks". Unlike classical approaches which associate intelligence with the explicit manipulation of structured symbolic expressions, this new approach proposes that intelligence emerges as a result of the transmission of activation levels within large networks of highly interconnected nodes, at times without specific semantic content. Implicit in the structure and behavior of neural models, they exhibit many of the interesting phenomena that are commonly associated with intelligence but which have eluded successful explication using traditional symbol-processing techniques. Research efforts over the past, however, have almost exclusively concentrated on low-level perceptual processing and/or pattern classification tasks. For high-level cognitive tasks which require knowledge-based decision-making such as does diagnostic reasoning, symbol processing approaches have continued to be the technique of choice.