Shallow Neural Networks for Channel Estimation in Multi-Antenna Systems

In this paper, we investigate neural network-based channel estimation strategies for point-to-point multi-input multi-output (MIMO) systems. In an attempt to keep computational complexity low, we restrict ourselves to shallow architectures with a single hidden layer. Specifically, we consider (i) fully-connected feedforward neural networks; and (ii) ID/2D convolutional neural networks. The analysis includes an assessment of the estimation error performance, along with the computational complexity as-sociated to the training and inference phases. Several benchmarks are considered, such as the conventional least squares or (linear) MMSE estimators, and other deep neural network architectures from the literature.