IEEE Wireless Communications Letters · 2021 · 36 citations · 11 references
ComputingComputer ScienceElectrical EngineeringDigital CommunicationMachine LearningWireless CommunicationsChannel EstimationDeep LearningInterference ManagementNeural NetworksTransfer LearningWireless Channel ModelingBit Error RateBit-error Rate AnalysisGeneralization ErrorInsufficient TrainingInterference EstimationMitigationRelay ChannelRelay Networks
Existing DNN‑based SI mitigation techniques require many online training samples, diverting resources from data communication. We propose and evaluate two new methods for digital mitigation of the non‑linear Self‑Interference (SI) signal in a Full‑Duplex relay link. The methods use DNNs to reconstruct non‑linear SI components, employ transfer learning and decouple SI signal reconstruction from channel estimation to reduce or eliminate online training, and are evaluated via link‑level simulation against state‑of‑the‑art DNN solutions. Simulations show the proposed techniques achieve robust BER even at SIR as low as –20 dB, and the authors provide an upper bound and approximation for the generalization‑error probability due to limited training data.
We propose and evaluate two new methods for digital mitigation of the non-linear Self-Interference (SI) signal in the context of a Full-Duplex (FD) relay link. The proposed methods utilize Deep Neural Networks (DNN) for the reconstruction of the non-linear SI components. Existing DNN-based SI mitigation techniques require a large number of on-line training samples, which takes resources away from actual data communication. The proposed methods allow a significant reduction in or complete elimination of on-line training by utilizing transfer learning and decoupling the reconstruction of non-linear transmitted SI signal from the estimation of the SI propagation channel, respectively. We utilize link-level simulation to demonstrate the SI cancellation performance of the proposed approaches compared to the existing state-of-the-art DNN-based solutions. The simulation results confirm that the proposed SI mitigation techniques achieve robust Bit Error Rate (BER) performance for Signal-to-Interference-Ratio (SIR) levels as low as −20dB. We derive an upper bound and an approximation for the probability of generalization error that is caused by an insufficient training set size.
11
On the Number of Linear Regions of Deep Neural Networks
Guido Montúfar, Razvan Pascanu, Kyunghyun Cho et al. · arXiv (Cornell University) · 2014 · 488 citations · Full text