Rule Extraction from Trained Artificial Neural Networks
Geczy, P. and Usui, S. (1999)
Behaviormetrika, Vol. 26 (1), pp. 89-106
- Nominator's statement
- Knowledge refers to the body of accumulated and stored information about a subject. It is believed that the brain is the medium for accumulating and processing information. The mature brain is able to process sensory information into coherent patterns of activity that form the basis of our perceptions, thoughts, and actions. Each function of a mature nervous system, from a simple reflex response to a complex behavior depends on the action of distinct neuronal circuits being adapted functionally and connectionistically by experience and learning. Evidence of learning and localization of various higher functions played an important role in connectionistic modeling. Connectionistic models resemble the brain in two main aspects. First, knowledge is acquired through a learning process. Second, interneuron connection strengths, known as synaptic weights, are used to store the knowledge. The resemblance of the connectionistic systems to biological neuronal circuits give rise to questions regarding knowledge representation and the possibility of knowledge extraction. Acquisition of knowledge from connectionistic models and its representation by logical formalism such as rules have an indispensable value in knowledge-based systems. Logical rules underline the expert knowledge about the system and can be used i for reasoning, explanatory purposes, and/or decision making. Last, but not least, the problem of knowledge extraction can shed light not only on our logical abilities, but also on their limitations. "Geczy and Usui propose a method of rule extraction from arbitrarily trained feed-forward neural networks. The proposed method does not impose any requirement (or constraints) on the network before training, and may have wider applicability than existing methods. The method is established on firm theoretical grounds that give conditions for equivalence of rules from a network and crisp logical formalism."
Nominations have closed, but you can still add a comment to this
nomination.
