Engineering and Digital Sciences

Energy consumption at the heart of research

From a latest-generation engine with an eco-responsible design to the development of artificial intelligence software, reducing energy consumption is central to engineering and digital innovation and raises questions about its applications.


Synapses and artificial neurons: the circuits of the future

© Sören Boyn / Unité mixte de physique CNRS/Thales

The AlphaGo software is capable of defeating champions of the game of Go and has marked artificial intelligence research in recent years. “It is a victory by the machine,” admits Laurent Larger, director of the Franche-Comté Electronics Mechanics Thermal Science and Optics – Sciences and Technologies (FEMTO-ST) Institute1CNRS/Université de Franche-Comté/ENSMM/Université de Technologie Belfort Montbéliard, “but a loss in terms of energy, with a consumption up to 10,000 times greater than that of the human brain.”
These AIs suffer from the limits of conventional computers, whose architecture has remained unchanged since the middle of the 20th century.

Whilst these are excellent for arithmetic calculations, they are totally ill-adapted to energetically efficient learning tasks. The FEMTO-ST Institute has thus chosen to prioritize photonic reservoir computing. This system reproduces neuronal organization in architectures that distribute information in time, using time-delay reservoirs, or in space, in the plane of a light beam. Ultra-rapid optical telecommunication components now make it possible to recognize nearly a million words per second.

Researchers from the Institute of Electronics, Microelectronics and Nanotechnology (IEMN)2CNRS/Université de Lille 1/ISEN Lille/Université de Valenciennes et du Hainaut-Cambrésis/Centrale Lille and the IRCICA3Institut de Recherche sur les Composants logiciels et matériels pour l’Information et la Communication Avancée (CNRS/Université de Lille) have designed artificial neurons that are a thousand times more rapid and energy-efficient than their biological equivalents. They function using a technology of conventional integrated circuits whose architecture has been completely rethought.

“A computer needs to be programmed, whereas a biological network learns for itself,” explains Alain Cappy, professor at the Université de Lille and researcher at the IRCICA. “Using artificial circuits that are much more rapid, we aim to reproduce what an infant takes years to achieve.”

Other teams, centred around the UMPhy joint physics unit4CNRS/Thales, the Centre for Nanoscience and Nanotechnology (C2N)5CNRS/Université Paris-Sud, and the Integration: from Materials to Systems laboratory (IMS)6CNRS/Université de Bordeaux/Bordeaux INP presented the first nano-neuron dedicated to vocal recognition, which identifies numbers with a 99.6 % success rate. Synthetic synapses were also created to link the artificial neurons, whilst adapting their resistance to the transmitted action potential. “We rely on all the physics of the components designed by Julie Grollier’s team7Researcher at the CNRS/Thales joint physics unit, rather than using a transistor merely as a switch or an amplifier,” explains Damien Querlioz, researcher at the C2N and recipient of the CNRS bronze medal. This approach enables a seemingly very simple electric neuron to carry out complex calculations.

Physical review X, February 2017.
Frontiers in Neuroscience, mars 2017.
Nature Communications, april 2017.
Nature, september 2017.

Sources :

Advances in the Ising Model

The mechanical statistics toolbox for studying the behaviour of molecules at their critical thresholds and beyond, is in constant evolution.

From rational argumentation to polysemy, computing improves its solutions

Beyond calculations, researchers help programs answer the most challenging requests.

Equations to fight tumours

To optimize the success of chemotherapy while preventing the appearance of resistant tumours, a mathematical model offers novel strategies.

From new materials to the study of impacts, aeronautics flies high

Although their use has become commonplace, airplanes nevertheless remain technological jewels that research constantly refines.

Holding the world in one hand

The reduced sphere is an excellent paradoxical geometrical object that maintains its surface and distances when its radius is reduced.

Light tuning for opto-mechanical nano-oscillators

Researchers managed to tune three distant nano-mechanical oscillators by using an infrared light to connect them.

Polymers to store data

Digital plastics have made amazing progress in their performance and are now capable of storing several octets of information.

Protein aggregation: a matter of irregular shapes?

Protein aggregation is causal in many neurodegenerative diseases, such as Alzheimer's. Protein aggregation in the form of fibres could arise from a mechanism that is not specific to these molecules.



In brIef

Exploring contents with ISTEX

ISTEX is a platform supported by the French government’s “Investments for the Future” programme, which provides access to multidisciplinary funding for the research and higher education community. ISTEX is integrated into the work environment of researchers and proposes knowledge mining tools to find answers to scientific questioning in all disciplines from a homogenized, enriched and accessible corpus.

QUAERO at Saclay

A facility has been built on the Saclay plateau, near Paris, to house the QUAERO French-German industrial research programme dedicated to automated analysis and the enrichment of digital, multimedia and multilingual content. As part of an ambitious project to set up a research centre of international scope for language science and technology, the building will facilitate interactions between researchers and students of the Université Paris-Saclay.