A study of the design and analysis of feed-forward neural networks

This thesis shows that a design and analysis system for feed forward neural networks is desirable, and that the currently available techniques do not work. Methods have been presented that solve the problem of analysis, showing that analysis is possible and desirable for classification networks. The biggest limitation is the size of the network and that the analysis tools are only applicable to properly designed classification systems. A method of reducing the size of classification networks is presented along with a design methodology for non classification systems.