Systems of feedforward ANNs

It is not difficult to find a problem that is very hard-if not impossible-to solve using a single feedforward network. This paper starts by describing such a problem and suggesting a solution. The solution takes the form the MetaNet system. MetaNet allows easy configuration of any topology of feedforward neural networks. The development of the MetaNet system is described and a specific implementation-simple analysis of harmony in western music is detailed. In drawing conclusions from the MetaNet system, an important limitation of static neural network system is identified: the lack of fluid integration of new knowledge. This limitation is addressed in a second system of feedforward neural networks-the FuseNet system. It allows for incremental increase of information contained in a system of networks. New information is integrated into networks without supervision while the complete system is functioning. The basic layout of the FuseNet system is described along with some initial results