Daniel S. Yeung, Hank-shun Fong
One of the major shortcomings of neural network as a problem solving tool lies in its opaque nature of knowledge representation and manipulation. For instance, the way that a learning algorithm modifies the connection weights of a network cannot be easily understood in the context of the application domain knowledge. Thus, the applications of neural networks is limited in areas where user’s understanding of the situation is critical. This paper introduces a facility called knowledge matrix for a rule induced Neocognitron network. It represents the correlation between the knowledge stored internally in the network and the symbolic knowledge used in the application domain. Another facility called response matrix is developed to represent the network’s response to an input. These two facilities are then employed cooperatively to generate symbolic interpretations of the network’s response. Based on the interpretations, queries can be made against the networks responses and explanations can be provided by the system. Two detailed examples are discussed. It can be shown that the network knowledge can be refined evolutionarily without degrading its comprehensibility. An algorithm has also been formulated to adapt the system with respect to one type of recognition error.