MIMO Full Duplex Radios with Deep Learning

Abstract

This paper presents a novel design for self-interference (SI) cancellation in multiple-input multiple-output (MIMO) full duplex radios by using a cascaded neural network structure. Our approach exploits the spatial correlation among the transmit and receive antennas in multi antenna systems to reduce the complexity of echo cancellation filters at the receivers. Specifically, we propose a method in which, instead of using a cancellation filter for each echo channel, we decompose it into a common filter for a single aggressor going into a number of victims each having their own secondary filters, thus, reducing the individual filter complexity. In fact, our method provides significant complexity reduction with minimal degradation in performance compared to the naive approach of replicating the single-input single-output (SISO) design into the MIMO system. We further explain the theoretical framework for functioning of neural network for echo cancellation in both SISO and MIMO systems. We show that using a one layer Rectified Linear unit (ReLU) neural network to solve the SISO SI problem is theoretically optimal. We evaluate our approach on a series of simulated correlated channels to show case complexity reduction.