Optimal Modular Feedfroward Neural Nets Based on Functional Network Architectures

Abstract

Functional networks combine both domain and data knowledge to develop optimal network architectures for some types of interesting problems. The topology of the network is obtained from qualitative domain knowledge, and data is used to fit the processing functions appearing in the network; these functions are supposed to be linear combinations of known functions from appropriate families. In this paper we showthat these functions can also be estimated using feedforward neural nets, making no assumption about the families of functions involved in the problem. The resulting models are optimal modular network architectures for the corresponding problems. Several examples from nonlinear time series prediction are used to illustrate the performance of these models when compared with standard functional and neural networks.KeywordsHide LayerFunctional NetworkModular NetworkFunctional FamilyModular Neural NetworkThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.