AI4CODE

Thanks to sixty years of continuing research and practice in coding theory, forward-error-correction (FEC) at medium to long block length has now reached a mature state. A fair amount of experience has been accumulated on the design of codes and implementation of message-passing decoders that either achieve or, at the very least, closely approach channel capacity at data rates of several tens or even hundreds of Gb/s. Still, many challenges remain ahead in the field of channel coding. Meanwhile, the success of deep learning and continuous improvements in computing capabilities over the past decade have fueled a resurgence of interest in applying machine learning (ML) techniques to a variety of communication problems. The availability of off-the-shelf learning software packages now makes it possible to parametrize communication systems and algorithms or even replace them by generic, black-box deep neural networks, and to train them to perform at least as well as the state-of-the-art. Application to code design and decoding is no exception to this general trend. Yet the application of ML to channel coding is still in its early stages, and its potential benefits for the field far from being sufficiently investigated and understood.

Objective of the project

The aim of the AI4CODE project is therefore to explore and assess how ML techniques can contribute to improvements in coding theory, techniques, and practice. The focus is placed on FEC, and the project is built around four inter-related objectives.

• Objective #1: Explore how ML can contribute to improving the state of the art in FEC decoding.
• Objective #2: Investigate how ML techniques can improve current knowledge and practice in FEC code design.
• Objective #3: Learn from the machine
• Objective #4: Develop a general expertise and critical thinking on ML algorithms and their applications to coding theory and practice

Employed Methods

Different kind of coding problems will be considered in AI4CODE. They are the result of a careful selection by the project members, based on their long-standing experience in the field. All are regarded as ideal playgrounds for experimenting with ML. Most of them are difficult in the sense that we lack a clear theoretical understanding so far. Any progress on either of these problems is expected to generate advances that could impact, in turn, the design of communication systems in the short term, by building intelligently on top of existing knowledge and practice. The selected coding problems can be grouped into four main research tracks:

• Track 1: Learn to correct better & smarter
• Track 2: Revisit design paradigms to arrive at better codes
• Track 3: Learn new families of codes
• Track 4: Pave the way towards decoders that automatically adapt to channel uncertainty

The central goal of the AI4CODE project is not to replace expert knowledge by black-box algorithms, but ultimately to learn from the machine. All project members have worldwide recognized expertise in the design and decoding of error-correcting codes. Our driving methodology will be to capitalize upon this expertise and to favor a model-based approach by augmenting our legacy code design methods and decoders with learning algorithms wherever relevant. If the machine comes up with new code designs or decoding strategies that outperform the state-of-the-art, then we will inspect and try to interpret the trained solutions in order to infer why they work better, with the intent of obtaining new theoretical hindsight that could ultimately translate into improved design tools or decoders.

Expected outcomes

Building upon the consortium worldwide-recognized long-standing experience in FEC, the outcomes of AI4CODE are expected to represent a significant step forward in the design and decoding of next-generation FEC codes. FEC code specification is an important part of the standardization process of communications systems. Members of the AI4CODE team have already taken active part towards the adoption of FEC solutions, Turbo codes for instance, in past worldwide standards. We note a global growing interest in applying Artificial Intelligence techniques, especially deep learning, to channel coding. The main drivers are coming for the most part either from Asia or from the US. Increasing our level of expertise and proficiency on these techniques is thus of prime importance to be able to evaluate the technical solutions proposed by others, and be in position to push forward competitive, proprietary alternatives in emerging standards.

IMT Atlantique Role

IMT Atlantique is the project leader, and will also be involved in each of the four research tracks listed above, in close collaboration with the other project partners. IMT Atlantique has long-standing expertise in FEC with an emphasis on sparse-graph codes, iterative decoders, and their hardware implementation. In particular IMT Atlantique is a world-wide recognized leader in Turbo code design and has actively contributed to their adoption in many communication standards.

Partners

The AI4CODE project brings together a team of six French research institutes with complementary, world-wide recognized expertise in FEC, and a long history of working together. LabSTICC/IMT Atlantique partners in this project are:

• CEA LETI, Grenoble, has wide research experience in the design of codes on graphs and message- passing decoders, with a particular focus and interest in LDPC and Polar codes, whether binary or non- binary
• IMS/Bordeaux INP, Bordeaux, UMR 5218, is concerned with Algorithm-Architecture-Matching for FEC software and hardware decoders in digital communication systems. IMS has investigated high-speed simulation of Polar Code with x86 & GPU architectures, and supports the AFF3CT open-source software toolbox which provides a high-performance simulation environment to evaluate FEC coding schemes.
• IRIT/Toulouse INP-ENSEEIHT, Toulouse, UMR 5505, has a research expertise in signal processing for digital communications, iterative receiver design, and channel coding at large, with key strengths in LDPC code design as well as in Polar code design and decoding
• ETIS/Cergy Paris Université (CYU), Cergy Pontoise, UMR 8051, has a strong background in the design and decoding of sparse-graph codes, both binary and non-binary, and their application to emerging networking and edge computing problems
• Lab-STICC/Université Bretagne Sud (UBS), Lorient, UMR 6285, has recognized experience in the design and decoding of error control codes, both from the algorithm and hardware architecture point of view, and a recent strong interest in non-binary codes and low-rate modulation for short-packet communication