Gradient-descent training for phase-based neurons

This paper offers details on a particular type of Complex Valued Neural Networks, which are Artificial Neural Networks that accept complex-valued inputs and use complex numbers for the values of the internal parameters. The functioning in the complex numbers domain grants CVNNs more computational power than classical ANNs. Phase-Based Neurons (PBNs) are simple CVNNs which use for the internal weights complex numbers with the modulus 1, the only adaptable parameters being the phases. We describe in this paper an improved method for PBNs training, showing its performance in learning linearly non-separable logical functions.