Learned CSI Feedback for Massive MIMO

Overview

In frequency division duplexing (FDD) massive MIMO systems, the base station cannot infer the downlink channel from the uplink. Each user estimates its channel from downlink pilots and sends channel state information (CSI) back to the base station, which uses it to compute precoders. As arrays grow, this feedback takes up more and more uplink resources. Most designs compress the CSI so that the base station can reconstruct the channel, even though the base station only needs it to choose precoders.

This project designs the feedback for the precoding task instead. A deep learning architecture learns the pilots, the users’ encoders and quantizers, and the base station’s precoding network jointly. Training uses a loss that maximizes the users’ sum rate while minimizing the number of feedback bits. In simulations, this precoding-oriented design reaches higher sum rates than reconstruction-oriented feedback, especially at low overhead. A journal version extends the design to multi-cell, multi-user MIMO systems.

The work is a collaboration with Sivarama Venkatesan, Jinfeng Du and Harish Viswanathan of Nokia Bell Labs.