Neural Networks with Complex-valued Neurons for Recurrent and Feedforward Architectures

                          Jacek M. Zurada
                      University of Louisville


This talk discusses neural networks with complex-valued neurons with both discrete 
and continuous outputs.  It reviews existing approaches for their use in fully coupled 
associative memories. Such memories are able to process multiple gray levels and can 
be applied for image de-noising.  In addition, when complex-valued neurons are 
generalized to assume a continuum of values, they can be used as versatile 
complex-valued substitutes for popular real-valued perceptron networks. Learning of 
such neurons is demonstrated and described in the context of traditional multilayer 
feedforward network learning.  Learning of complex-valued networks is derivative-free 
and it usually requires reduced network architecture. The notion of a universal binary 
neuron is also introduced. Selected examples and applications of such networks in 
bioinformatics and pattern recognition are also discussed.