Mapping rule-based systems into neural architecture

摘要

A novel approach has been developed that maps a rule-based expert system into the neural architecture in both the structural and behavioural aspects. Under this approach, the knowledge base and the inference engine are mapped into an entity are called ‘conceptualization’, where a node represents a concept and a link represents a relation between two concepts. A concept node is designated by a small number of language symbols. In the neural system transformed from a knowledge-based system, the inference behaviour is characterized by propagating and combining activations recursively through the network, and the learning behaviour is based upon a mechanism called ‘back-propagation’, which allows proper modification of connection strengths in order to adapt the system to the environment. This approach is based on the analogies observed between a belief network and a neural network, and its validity has been demonstrated by experiments. Finally, the advantages and disadvantages of this approach are discussed with respect to inference and learning.