ABSTRACT
In this chapter, we present the results of a study of a new version of the LAPART adaptive inferencing neural network [1], [2]. We will review the theoretical properties of this architecture, called LAP ART-2, showing it to converge in at most two passes through a fixed training set of inputs during learning, and showing that it does not suffer from template proliferation. Next, we will show how real-valued inputs to ART and LAPART class architectures are coded into special binary structures using a preprocessing architecture called Stacknet. Finally, we will present the results of a numerical study that gives insight into the generalization properties of the combined Stacknet/LAPART-2 system. This study shows that this architecture not only learns quickly, but maintains excellent generalization even for difficult problems.
