Residual Neural Networks for Learning the Full-Duplex Self-Interference

Full-duplex (FD) is a key technology for enhancing the capacity of next-generation wireless systems by jointly maximizing the utilization of time and frequency resources, resulting in low latency and high spectral efficiency. However, the self-interference (SI), leaking to the receiver chain from its own transmitter chain, is the main issue that hinders reaping the key benefits of FD systems, and SI cancellation (SIC) is introduced to enable such benefits. Digital non-linear SIC is traditionally performed using model-driven approaches, such as polynomial models, which are of high complexity. Thus, data-driven machine learning (ML) approaches are introduced for learning the FD non-linear SI with lower complexity. This paper proposes an ML approach based on residual neural network (Res-NN) to learn the FD non-linear SI and relax the computational requirements of the traditional methods. Res-NN uses shortcut connections from the input/hidden layer to the output layer to enhance the learning capabilities of the SI cancelers. Simulation results show that an NN employing residual connections could effectively learn the FD SI and outperform the existing benchmarks in the literature.