| dc.description.abstract | Machine learning is a highly promising tool to design the physical layer of wireless communication systems, but its scaling properties for this purpose have not been widely studied. Machine learning algorithms are typically evaluated to learn SISO communications and low modulation orders, whereas current wireless standards use MIMO and high-order modulation schemes to increase capacity. The memory requirements of current machine learning algorithms for wireless communications increase exponentially with the number of antennas and thus they cannot be used for advanced physical layers and massive MIMO. In this paper, we study the requirements of end-to-end machine learning models for large-scale MIMO systems, determine the bottlenecks of the architecture, and design different solutions that vastly reduce overhead and allow training higher MIMO and modulation orders. We show that by training the autoencoder in a bit-wise manner, the memory requirements are reduced by several orders of magnitude, which is a critical step for machine learning-based physical layer design in practical scenarios. Besides the reduced memory requirements, our design also improves performance over the classical autoencoder for MIMO systems. | es |
