20-000000856: Empirical Aspects of Inventing PHY layer Via Deep Learning
Abstract
Approved for Public Release:Deep learning has demonstrated success in a wide variety of domains, ranging from computer vision, speech/language processing, to games (eg: Go, Starcraft, chess), transforming our daily lives. The success can be attributed to either the model deficiency or algorithm deficiency; model deficiency refers to the lack of analytical models for the inference problems (inthe case of computer vision and natural language processing). Algorithmic deficiency refers to the lack of efficient algorithms dueto the enormous space of algorithms (in the case of games). In designing PHY layer for practical channels, we face both model and algorithmic deficiencies. The underlying channels often do not have a precise mathematical model, in which case current designs rely on heuristics. For example, when a jamming noise occurs, channel decoders first threshold the received values. The space of algorithms for PHY layer is huge; for example, a channel coding function of k bits require a mapping from 2^k messages to 2^k codewords (andk ranges from 10s to 1000s). Motivated by the success of deep learning, we posit that deep learning methods can play a crucial rolein the design of physical layer communication algorithms. Our central hypothesis is that we can learn (i) channel decoders that aremore reliable and robust than the heuristic decoders for practical and tactical channels and (ii) codes (both channel encoders and decoders) that outperform known codes for channels with feedback, a setting which lacks practical reliable codes. Our hypothesis is largely substantiated based on our extensive success in inventing channel decoders and codes via deep learning [1-5]. The goal of our first thrust is to invent channel decoders that are robust and adaptive to several real-world challenges (including hardware imperfections, channel estimation errors, interference) and potential attacks (jamming) and to demonstrate our invention in an over-the-air implementation. Our approach is to model the channel decoder as neural networks and learn the parameters of neural networks in a supervised training. The goal of our second thrust is to demonstrate that channel codes, along with channel decoders, learned via deep learning can significantly outperform the known codes for channels with feedback and utilize the feedback more efficiently than the known codes. Our approach is to model both the encoder and decoder by neural networks and train them jointly on simulated channels. We will apply the know-hows from substantial experience in neural network training [1-5], compression[6-8], harmoniously with information theoretic insights. In Thrust 1, we will deliver a neural decoder library which includes neural network based decoders trained to be robust and adaptive to several practical channel models with impairments for turbo and LDPC codes. We will also verify theperformance of neural decoders via over-the-air experiments. We expect 2-10$times$ improvementin bit error rates (BER), depending on the Signal-to-Noise ratio (SNR) and encoder/channel parameters, with the latency comparable to the latency of traditional decoders. We will also characterize the trade-off between the reliability and latency. In Thrust 2, we create a Deepcode library which includes neural network based encoders and decoders designed for various conditions of channels with feedback and optimized for its latency. We expect 2-10 improvement in BER upon the state-of-the-art depending on the channels.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 05, 2021
- Source ID
- N000142112379
Entities
People
- Hyeji Kim
Organizations
- Office of Naval Research
- United States Navy
- University of Texas at Austin