A KNOWLEDGE MATRIX REPRESENTATION FOR A RULE-MAPPED NEURAL-NETWORK

Neural networks are noted for their learning and generalizing capabili ties. However, their advancement and applicabilities are severely limi ted by the low comprehensibility of their internal knowledge. Previous ly, the authors have proposed a rule-mapped neural network model, by i ncorporating domain knowledge initially. This paper suggests a tool na med 'knowledge matrix', that produces symbolic interpretations of the network's response to an input. These interpretations can be shown to enhance the reasoning power of the system. Moreover, the system knowle dge can be refined explicably. The proposed approach is tested with a Chinese character structure recognition problem. This is an attempt to model a human learning process that is commonly observed in many situ ations. For example, a trainee may be given a set of provisional rules at the very beginning which is expected to be moderated in accordance with his forthcoming experience.