Lecturer at IMT Atlantique and member of the Lab-STICC laboratory, Elsa Dupraz has been awarded the Prix Espoir IMT-Académie des sciences for her research into error correcting codes, which lower power consumption of communication systems. A field with a large range of applications.
You are the 2024 winner of the Prix Espoir IMT-Académie des sciences, for your work on error correcting codes. What is the principle behind these codes?
In all digital communication systems - telecoms networks, wifi, etc. - there are hazards: background noise, reverberation, distortion, etc. They disrupt transmission and generate errors. Correcting codes (also known as “channel coding”) are used to compensate for these errors. How do they work? Essentially, by introducing redundancies into the message. Even if the message is longer, this reduces the transmission power of the signal, which helps saving a lot of energy. Nowadays, reducing the power consumption of digital devices is a priority. But that is not all: if a device consumes more power, its battery weakens more quickly, and it loses autonomy... That is why we are looking to design codes that are both “simple” enough to consume little energy, and that guarantee reliable transmission.
These correcting codes are now widespread. They are used in every transmission systems: telephone, computer, wifi, but also for data compression.
Elsa Dupraz and Françoise Combes, Vice-President of the Académie des Sciences
Photo credit © Académie des sciences - Mathieu Baumer.
What amount of energy can be saved thanks to these codes?
It varies tremendously from case to case. By combining various methods, you can achieve a gain of 50% - and sometimes much more.
What methods do you specifically use?
With my team of four PhD students, one post-doctoral student and various collaborators in France and abroad, we are working on three different avenues. The first is DNA storage - an idea dating back some ten years. Data storage consumes a lot of energy, as we can see with data centers. DNA, on the other hand, has two advantages: it is very dense, because its molecules are so small - some say that all the world's data would fit in a small DNA van. And it is robust: it lasts a long time (several decades at least) and can withstand high and low temperatures.
How do you store information in DNA molecules? It is not that easy, because you have to go from digital to living. In other words, we have to convert binary files into quaternary files, in accordance with the structure of DNA (with its four key elements, the bases A, T, C and G). To do this, we use inactivated synthetic DNA, which biologists know how to produce for medical use. In reverse, the data stored in the DNA is read, using the same procedure as for sequencing.
But reading the information stored this way is not always reliable. It contains errors, and these are different from those encountered in telecoms: there is no noise, but rather substitutions, deletions or insertions (additions of bases). My job is to detect these errors, and to develop correcting codes designed for DNA, using dedicated algorithms. This project is part of the major PEPR MoleculArXiv(1), which brings together about twenty French research teams, with funding of 20 million euros until 2029. A channel simulator we designed, available as open-source software, is already being used by several PEPR teams.
What is the second avenue you are working on?
It involves implementing computing architectures in memory. A processor is generally made of two parts: a computing system and a memory. Between the two, incessant movements inevitably consumes a great deal of energy. This is particularly true of AI, with its neural networks containing an extremely high number of parameters. To avoid these multiple paths, it is possible to combine the computing system and the memory. This involves using new types of memory, known as non-volatile random access memory (NVRAM). The problem is that they also produce errors... and that it can then have an impact on the calculations!
Therefore, our work is aimed at designing codes capable of recovering the correct result - if possible without adding too many redundancies to the calculation. We have developed a new particular method for predicting the effect of faults in neural network weights. We also use special decoders, known as LDPCs (Low Parity Density Codes), which are robust to faults. Thanks to these tools, AI methods could eventually be integrated into in-memory computing architectures.
And what about your third research topic?
Il concerne l’apprentissage sur données compressées. Pour entraîner un modèle d’IA, il faut en effet It is about learning on compressed data. To train an AI model, you need to provide it with huge quantities of data - millions of images, for example - which are usually compressed. Decompressing this data is both time-consuming and energy-intensive. One solution is to dispense with decompression, and learn directly from the compressed data, without having to reconstruct it.
But conventional compression tools (“entropy” coders) are not suitable, as they do not preserve certain data localization properties. Hence our team's idea to redesign the compression chain, replacing some of the compression tools with “linear codes”, derived from channel coding. This process improves learning speed while maintaining efficient compression. Significant reductions in energy consumption can be expected. This project has received funding from the Labex CominLabs, dedicated to digital technologies. In addition, next year we plan to develop a Jpeg image encoder software demonstrator, which will be open to the research community.
How do you manage to work on three such different projects at the same time?
My specialty is correcting codes. It is a field that is common to all three projects I'm working on. This means that I have to work with a wide variety of teams: biologists and bioinformaticians for DNA storage, specialists in information theory and compression for learning from compressed data, experts in hardware architectures for in-memory computing... You have to understand the logic and problems of each of them, while contributing your own expertise. It is a highly multidisciplinary approach, and a stimulating exercise.
Which of these three avenues do you see as the most promising?
In the short term, compression undoubtedly offers the best prospects. In-memory computing is more of a long-term proposition. As for DNA storage, it is still a gamble for the rather distant future - especially as the high cost of synthesizing DNA molecules is a brake. Microsoft has, however, presented a demonstrator in 2019.
Many players have interest in all this work, but it is still in the early stages of research. Moreover, in the field of error correcting codes, major challenges remain. For example, we do not really know how to proceed with very short messages - such as those emitted by sensors used in industrial processes. Also, as far as I am concerned, error correcting codes are not limited to telecoms: they can be applied in many other fields, such as those we have just discussed, but also in the factory of the future or the autonomous car.
What does this “Espoir” award mean to you?
It is, of course, a true recognition of my work and my team at IMT Atlantique, and an encouragement to continue in this direction. It is also a way of drawing attention to the field of corrective coding. It is a generic tool that can help solve a number of important problems. Also, the award confirms Brest's status as one of France's and Europe's leading coding centers. A modern coding technology, “turbo-codes”, was invented in Brest in the 1990s by researchers at Télécom Bretagne, which became IMT Atlantique in 2017.
- (1) Priority research programs and equipment.
Elsa Dupraz is congratulated by Cécile Dubarry, CEO of the IMT, François Baccelli, President of the jury for the IMT-Académie des Sciences prize, and Yann Busnel, Scientific Director of the IMT.
Elsa Dupraz's career
“Oddly enough, most of the institutions I attended have merged with others and changed their names,” Elsa Dupraz laughs. The young researcher first studied at the electricity-electronics department of ENS Cachan, now ENS Paris-Saclay, then at Supélec (now Centrale-Supélec). She completed her thesis (on data compression) at the Signals and Systems Laboratory (LSS) at the Université Paris-Sud, now known as Paris-Saclay. This was followed by a post-doctoral program at the University of Arizona in Tucson, in conjunction with the Université Cergy-Pontoise (now CY Cergy Paris Université). In 2015, she started working as an associate professor at Telecom Bretagne, merged in 2017 with Mines Nantes to become IMT Atlantique... Since 2023, Elsa Dupraz has also held an HDR (Habilitation to Direct Research).
https://elsa-dupraz.fr/
Prix IMT-Académie des sciences
Since 2017, the Institut Mines-Télecom (via its Foundation) and the Académie des Sciences have awarded every yeat a “Grand Prix” and a “Prix Espoir”, the latter for scientists under the age of 40. The two prizes are awarded to researchers who have made a particular contribution to “advancing issues arising in industry or business, in the service of a sustainable economy”. Four fields are concerned: the industry of the future, digital sovereignty and sobriety, energy and environmental engineering and health.
by Pierre-Hervé VAILLANT
