A potential replacement for the conventional neuron is introduced. This is called a Macronet element and uses multiple channels per signal path with each channel containing two trainable non-linear structures in addition to a conventional weight. The authors show that such an architecture provides a rich spectrum of higher order powers and cross products of the inputs using less weights than earlier higher order networks. This lower number of weights does not compromise the ability of the Macronet element to generalise . Results from training a Macronet element to develop a relationship from a sparse map of Europe are given.
Lecture notes in computer science : developments in applied artificial intelligence : Proceedings of the 15th International Conference on Industrial and Engineering Applications of Artificial Intelligence and Expert Systems (IEA/AIE 2002), Cairns, Queensland, Australia, 17-20 June 2002,
Vol. 2358 (2002), pp. 103-119