Complex backpropagation neural network using elementary transcendental activation functions

摘要: Designing a neural network (NN) for processing complex signals is challenging task due to the lack of bounded and differentiable nonlinear activation functions in entire domain C. To avoid this difficulty, 'splitting', i.e., using uncoupled real sigmoidal imaginary components has been traditional approach, number fully introduced can only correct magnitude distortion but not handle phase distortion. We have previously NN that uses hyperbolic tangent function defined showed most practical signal problems, it sufficient an almost everywhere domain. In paper, design extended employ other hyperbolic, circular, their inverse family. They are shown successfully restore amplitude distortions non-constant modulus modulated signals.