Rule Refinement and Extraction with Local-Function Neural Networks

Robert Andrews, Queensland University of Technology

10:00 am Tue. Apr. 9 in 2310 CS&S

This talk presents an overview of rule extraction and rule refinement techniques that have been developed specifically for use with local function networks. Local function networks are artificial neural networks that make use of some form of local response units in their hidden layer. Networks that fall into this category include Radial Basis Function (RBF) networks, the Rapid BackProp (RBP) network, and the Rectangular Basis Function (RecBF) network. Techniques to be discussed include those described by Tresp, Hollatz, and Ahmad for rule extraction/refinement from RBF networks, Berthold and Huber Andrews and Geva for rule extraction/refinement from RBP networks. The seminar also discusses advantages and disadvantages of local solutions.