Artificial neural network with complex weight and its training
Y.-C. Shin, Rishika Iytha Sridhar · 1992
Artificial neural networks that use complex weights for the synaptic connections are presented. It is shown that the use of complex weights overcomes linear nonseparability for functions such as exclusive-OR and hence can be implemented using a single-layer network. The authors also present a modification to the backpropagation method to train the neural network presented. Several examples including symmetry problems, summation, and negation are presented to demonstrate the effectiveness of the use of complex weights. It is expected that this approach can implement functions of greater complexity using simpler networks (with fewer layers) than would be required with conventional approaches.>