Adaptive Data Augmentation for Deep Receivers

Deep neural networks (DNNs) allow digital receivers to learn to operate in complex environments. To do so, DNNs should preferably be trained using large labeled data sets with a similar statistical relationship as the one under which they are to infer. For DNN-aided receivers, obtaining labeled data conventionally involves pilot signalling at the cost of reduced spectral efficiency, typically resulting in access to limited data sets. In this paper, we study how one can enrich a small set of labeled data into a larger data set for training deep receivers without transmitting more pilots. Motivated by the widespread use of data augmentation techniques for enriching visual and text data, we propose a dedicated augmentation scheme for exploiting the characteristics of digital communication data. We identify the key considerations in data augmentations for deep receivers as the need for domain orientation, class (constellation) diversity, and low complexity. Our method models each symbols class as Gaussian, using the available data to estimate its moments, while possibly leveraging data corresponding to related statistical models, e.g., past channel realizations, to improve the estimate. The estimated clusters are used to enrich the data set by generating new samples used for training the DNN. The superiority of our approach is numerically evaluated for training a deep receiver on a linear and non-linear synthetic channels, as well as a COST 2100 channel. We show that our augmentation allows DNN-aided receivers to achieve gain of up to 3dB in bit error rate, compared to regular non-augmented training.