Multiuser Detection With Neural Feature Learning

Deep learning techniques demonstrate transformative progress in addressing structured non-linear decisions, providing a promising path to innovate engineering solutions. However, integrating learning modules with engineering design faces nontrivial challenges due to the fundamentally distinct practices. In this paper, we address such issues by connecting feature representation learning with the statistical dependence behind data, where we discuss the designs of training data, objective functions, and neural network architectures. As an illustrating example, we consider the multiuser detection problem in fading channels and present the receiver design by feature learning. In particular, we formulate the detection problem as learning the conditional dependence between the received and transmitted signals, conditioned on the channel state information. With this formulation, we design a deep neural network based receiver that adapts to various fading conditions without using online training samples. We also illustrate the incorporation of engineering domain knowledge in the network design and demonstrate its effectiveness by numerical experiments. Finally, we summarize several key principles in designing general learning-based engineering solutions.