Complexity-Scalable Neural Network Based MIMO Detection With Learnable Weight Scaling

This paper introduces a framework for systematic complexity scaling of deep neural network (DNN) based MIMO detectors. The model uses a fraction of the DNN inputs by scaling their values through weights that follow monotonically non-increasing functions. This allows for weight scaling across and within the different DNN layers in order to achieve scalable complexity-accuracy results. To reduce complexity further, we introduce a regularization constraint on the layer weights such that, at inference, parts (or the entirety) of network layers can be removed with minimal impact on the detection accuracy. We also introduce trainable weight-scaling functions for increased robustness to changes in the activation patterns and a further improvement in the detection accuracy at the same inference complexity. Numerical results show that our approach is 10 and 100-fold less complex than classical approaches based on semi-definite relaxation and ML detection, respectively.