Modification of backpropagation networks for complex-valued signal processing in frequency domain

Abstract

It is pointed out that existing backpropagation networks cannot be used for many signal-processing tasks because the signals are often represented with complex values. A complex-valued learning algorithm is derived for a backpropagation network where one or more of the inputs, outputs, and weight factors are complex-valued. The complex valued backpropagation network is tested with exclusive-OR problems for complex inputs, outputs, and weights. The network shows excellent mapping performance for this problem