2019 conference paper
Detection and Channel Equalization with Deep Learning for Low Resolution MIMO Systems
Conference Record - Asilomar Conference on Signals, Systems and Computers, 2018-October, 1836–1840.
Contributors: A. Klautau*, N. Gonzalez-Prelcic *, A. Mezghani * & R. Heath *
Deep learning (DL) provides a framework for designing new communication systems that embrace practical impairments. In this paper, we present an exploration of DL as applied to design the physical layer for MIMO systems with low resolution analog-to-digital converters. The application of DL is nontrivial thanks to the severe nonlinear distortion caused by quantization and the large dimensional MIMO channel. We investigate network architectures for channel estimation and detection. The channel estimation results indicate that the adopted DL architectures lead to good results in the large signal-to-noise ratio (SNR) regime, but are outperformed by state-of-the-art iterative message passing algorithms. For decoding, we adopted a multilabel classification architecture with implicit equalization and output size scaling linearly with the number of data symbols to be estimated. While feasible for high MIMO dimensions, the adopted DL architecture for decoding converged only for relatively small MIMO dimensions. A main conclusion of our paper is that DL still has potential but more efficient architectures are required, given the convergence problems associated with time-varying channels and 1-bit quantization.