Complex Valued Deep Neural Networks

Resource allocations in wireless communications play a major role in
designing efficient communication systems. To that end, research
interests have been devoted for developing resource allocations
schemes adopting convex optimisation tools. Those schemes offer
optimal solutions to the problems. However, they either require
computational processing capacity or are times consuming which does
not fit for practical systems. Deep Neural Networks (DNN) can be used
in those cases to approximate the optimal resource allocation schemes.

Since wireless resource allocation approaches often deal with complex
valued data, the adoption of DNN to the field of wireless
communications is not trivial. This is due to the fact that most of
the DNNs are dealing with real valued data. I have tried to convert
complex valued data into double-dimensional real data to fit with real
valued DNN. However, the performance of such systems in terms of
accuracy is not good. I would like to learn if there is an effective
approach to deal with complex valued data for DNN.