A metalevel architecture for knowledge-based neural network design

Abstract

In this paper, the term knowledge-based neural network (NN) design is used to refer to all efforts at exploiting prior knowledge in neural network configuration and training. A variety of techniques have been proposed for this purpose; SCANDAL provides a workbench for evaluating and integrating these techniques. After a quick overview of three main approaches to NN design, we describe SCANDALS multi-agent, metalevel architecture as well as its strategies for maximizing the use of domain knowledge. To assess the impact of prior knowledge on NN performance, experiments were conducted comparing knowledge-based techniques with a search-based configuration algorithm. Results show that the use of prior knowledge in neural network design leads to both faster learning and improved generalization. More interestingly, this appears to hold even when domain knowledge and data are deficient; in such cases, knowledge is extracted from the available data and is used both to configure the network and to generate artificial training instances. This leads us to hope that time-consuming iterative search can be avoided even in knowledge-lean domains.