Synthesizing neural networks for pattern recognition

It is proved that a three-layered neural network for pattern recognition can be synthesized if a pattern is separated by hyperplanes into a single region, and, if not, a four-layered neural network is synthesized. Then, it is shown that an n-input parity circuit is synthesized by a three-layered neural network with n hidden neurons. Finally, it is demonstrated how the recognition rate is improved by tuning the weights of the network generated by the backpropagation algorithm. As examples, a parity circuit was synthesized by a three-layered neural network and, for number recognition, the recognition rate was drastically improved by tuning the weights of the network