We assess the relative merits of a number of techniques designed to determine the relative salience of the elements of a feature set with respect to their ability to predict a category outcome-for example, which features of a character contribute most to accurate character recognition. A number of different neural-net-based techniques have been proposed (by us and others) in addition to a standard statistical technique, and we add a technique based on inductively generated decision trees. The salience of the features that compose a proposed set is an important problem to solve efficiently and effectively, not only for neural computing technology but also in order to provide a sound basis for any attempt to design an optimal computational system. The focus of this study is the efficiency and the effectiveness with which high-salience subsets of features can be identified in the context of ill-understood and potentially noisy real-world data. Our two simple approaches, weight clamping using a neural network and feature ranking using a decision tree, generally provide a good, consistent ordering of features. In addition, linear correlation often works well.
In this paper we address the problem of constructing reliable neural-net implementations, given the assumption that any particular implementation will not be totally correct. The approach taken in this paper is to organize the inevitable errors so as to minimize their impact in the context of a multiversion system, i.e., the system functionality is reproduced in multiple versions, which together will constitute the neural-net system. The unique characteristics of neural computing are exploited in order to engineer reliable systems in the form of diverse, multiversion systems that are used together with a "decision strategy" (such as majority vote). Theoretical notions of "methodological diversity" contributing to the improvement of system performance are implemented and tested. An important aspect of the engineering of an optimal system is to overproduce the components and then choose an optimal subset. Three general techniques for choosing final system components are implemented and evaluated. Several different approaches to the effective engineering of complex multiversion systems designs are realized and evaluated to determine overall reliability as well as reliability of the overall system in comparison to the lesser reliability of component substructures.