Investigators
Z. Obradovic, J. Fletcher, S. Milenkovic, P. Romero and R. Srikumar
Problem
Traditional neural network learning involves optimization of the interconnection weights between neurons on a pre-specified networkarchitecture. Determination of an appropriate architecture is a challenging problem which is typically approached through an expensive trial-and-error process.
Results
As a more feasible alternative for large scale applications, this study proposes several efficient parallel learning algorithms that simultaneously determines an appropriate number of neurons and their interaction parameters in a problem specific manner (srikumarainpress), (srikumarbinpress), (fletcher96j), (milenkovic96a), (fletcher94), (srikumar94), (fletcher93a), (milenkovic96), (milenkovic93). Experimental results indicate that our algorithms automatically construct near-minimal architectures as desired for good generalization. Using our constructive learning algorithm, a pre-existing rule-based system is integrated with information extracted from examples resulting in improved prediction quality on a financial advising problem (romerobook), (fletcher93j), (romero95), (fletcher93b).