On the design and initialization of layered feed-forward neural networks

This paper considers the design and initialization of a network, based on application specific knowledge, available at design time. We describe a methodology for translating high level knowledge about an application into a neural network interconnection specification. The program's design philosophy stresses separation of neural design from network function, a uniform syntax for neurons, inputs, and outputs, and flexibility in modularizing the resulting network. The ability to train neural networks allows the encoded knowledge to be further fine tuned for a specific data space. Moreover, the translation rules can allow for the selective training of subnetworks.