A FRAMEWORK FOR-IMPROVED TRAINING OF SIGMA-PI NETWORKS

This paper proposes and demonstrates a framework for Sigma-Pi networks such that the combinatorial increase in product terms is avoided. Thi s is achieved by only implementing a subset of the possible product te rms (sub-net Sigma-Pi), Application of a dynamic weight pruning algori thm enables redundant weights to be removed and replaced during the le arning process, hence permitting access to a larger weight space than employed at network initialization, More than one learning rate is app lied to ensure that the inclusion of higher order descriptors does not result in over description of the training set (memorization), The ap plication of such a framework is tested using a problem requiring Sign ificant generalization ability. Performance of the resulting sub-net S igma-Pi network is compared to that returned by optimal multi-layer pe rceptrons and general Sigma-Pi solutions.