Engineering reliable neural networks

The notion of multiversion system design is imported from software engineering where it has sometimes been used as part of a strategy for producing highly reliable software. We have further developed and refined this notion such that we can confidently undertake to improve the performance of any single neural network. For a number of reasons neural computing is better suited for use with a multiversion strategy than the conventional computing from whence the basic idea came. We have developed a methodology to underpin a multiversion approach to highly reliable neural net implementations. We present this methodology and several different applications of it (e.g., single level and two-level multiversion systems) that demonstrate the generalisation improvements obtainable within the general framework of a diverse, multiversion approach. A variety of results are compared and contrasted. They indicate that significant generalisation improvements can be obtained by a variety of different means.