Deep Learning for Channel Coding

Abstract

Wireless Communication has become a critical backbone of the information economy in the past few decades. In this rapidly improving telecommunications landscape, a crucial role is played by channel codes. Channel coding refers to the coding of information in such a way that the transmission can be robustly decoded even under noisy conditions. Progress in channel coding research has been powered by sophisticated mathematics and driven solely by human ingenuity, and therefore, progress is necessarily sporadic. For example, from convolutional code (2G) to Polar code (5G), it took several decades of research efforts to develop a new generation of channel codes. Deep learning has revolutionized a wide variety of fields and modern AI systems built using deep learning techniques are now able to surpass humans as well as human-designed algorithms. Motivated by the success of deep learning in other fields, in this thesis, we study the role of deep learning in tackling telecommunication system design. The first part of this thesis shows that three major channel coding problems: (a) decoder design, (b) code design, and (c) feedback code design, can be automated by applying end-to-end deep supervised learning with near-optimal performance. We show surprisingly in several of these scenarios that even when the communication channels follow well studied and canonical (text-book) models, there is a significant performance improvement from deep-learning. This can be attributed not only to the ability of the learning-based methods to adapt to the channel statistics (because it is a well-known channel), but also to its ability to design sophisticated non-linear algorithms for both encoding and decoding. The second part of this thesis studies other learning paradigms such as meta-learning and federated learning (FL), which can be applied to channel coding problems to further demonstrate the versatility of neural networks.Meta learning based neural decoder shows significant efficiency on data and computation, compared to naive fine-tuning. Finally, inspired by the strong algorithmic connection between FL personalization and meta learning, we propose a personalized FL algorithm which improves personalization significantly.