Machine learning: how to combine learning efficiency and digital sobriety

Benjamin Guedj’s British adventure began in 2018, when he was a researcher in the Modal team at the Inria Lille-Nord Europe Centre. “Ever since my PhD, completed at Sorbonne University and Télécom ParisTech between 2011 and 2013, I have been interested in machine learning algorithms as the basic building blocks of artificial intelligence and how we can characterise and improve their statistical performance”, he explains. “In particular, my aim is to allow them to “generalise” their learning from data and use as few computational resources as possible, an approach that will help develop less energy-intensive digital technology. After joining Inria in 2014, I contributed to strengthening the mathematical foundations of machine of machine learning using  “PAC-Bayesian” theory”.

It was in response to an invitation by John Shawe-Taylor, one of the world's leading experts on this theory, that Benjamin Guedj joined University College London (UCL) during a sabbatical from Inria to work with British researchers on this subject. “I was in search of new horizons and discoveries, both scientific and geographical, and at UCL I was immediately welcomed into the community of artificial intelligence researchers and quickly took on many scientific and institutional responsibilities”, he recalls.

The young researcher's commitment his research interests and his formidable drive to build ties between the two sides of the Channel led to the creation of the Inria London Programme just a few months after his arrival in the UK, as well as the roadmap of Genesis, his scientific research team focusing on the notion of “generalisation” in machine learning, based in particular on PAC-Bayesian theory.

Artificial intelligence and its flagship component, machine learning, have seen spectacular growth in recent years, with applications in many fields. Tomorrow's personalised and predictive medicine, driverless vehicles and rational agronomy are all promising applications.

Artificial intelligence (AI) algorithms are developed to teach a machine to perform a task without being explicitly programmed to do so”, explains Benjamin Guedj. “AI holds many promises for the future, some of which - such as that of autonomous synthetic intelligence - seem to be greatly overestimated! The solutions provided by these techniques are currently more in the field of augmented intelligence, giving humans tools to help them make decisions in complex and uncertain environments by making it possible to analyse a variety of situations and take account of numerous parameters, at a largely impossible speed for a human brain

Through machine learning, a computer program learns to perform a given task such as image recognition. “For a machine to be able to recognise a cat in a photograph with an efficiency close to 100%, the algorithm needs to analyse hundreds of thousands, even millions of images of cats, which have to have been collected beforehand and checked to make sure that they really do show cats!” explains Benjamin Guedj. The machine’s ability is based on its computing power, which is particularly energy-intensive and is paradoxically limited by this “brute force” approach.

“The learning capacity of humans is still incomparably more efficient than that of machines: a child recognises a cat after seeing one for the first time and is capable of transferring this knowledge to understand that a dog or a rabbit, like a cat, is an animal, because they all share common features (ears, fur, legs, etc.)”, says Benjamin Guedj.

It is this “ability to generalise” that interests the Genesis researchers. They aim to give machines the same capacity based on the progress of PAC-Bayesian theory, which allows forecasts to be made with minimum error and high probability (known as the “probably approximately correct”, or PAC theory) while incorporating a risk model (known as the “Bayesian” approach). This method of modeling uncertainties in machine learning processes opens the door to systematic analysis of the “generalisation” property.

This research programme also aims to develop efficient algorithms, allowing complex tasks to be carried out at lower computational costs. It is a way of contributing to building a frugal digital world. “For a simple cognitive task such as recognising an element in a picture, it is estimated that the energy consumed by a machine exceeds that of a human brain by 4 to 6 orders of magnitude! In a context where the financial and environmental cost of energy production is rising significantly, improving the energy efficiency of AI is a necessity and a major scientific challenge”, explains Benjamin Guedj.

The researcher and his colleagues are starting to obtain promising results in the fields they are exploring, as shown by the numerous publications already shared with the scientific community. Benjamin Guedj's commitment to the project goes beyond his scientific role and the researcher wants to help make Inria London a hub for the Institute in the UK. It is a way of taking concrete action to bring scientists from the two countries together on the strategic theme of artificial intelligence.  

2022 may be one of the most memorable years in Benjamin Guedj’s already promising career. Aged 35, he recently joined the ranks of the 2022 cohort of Young Leaders of the Franco-British Council and in the coming months will be awarded the prestigious title of Knight of the Ordre des Palmes Académiques during a ceremony to be held at the French Embassy in the UK.

These two distinctions recognise the young Inria researcher’s commitment to developing major scientific collaborations between France and the UK, particularly in machine learning. Benjamin Guedj was at the origin of a partnership with University College London in the field of artificial intelligence (AI), which was given concrete form in 2020 with the creation of the Inria London Programme, Inria’s national programme, and the creation of a joint Inria-UCL project team, Genesis, composed of more than 40 people (including six permanent staff).

• « 15 Minute Meets » with Benjamin Guedj, UCL Centre for Artificial Intelligence, 10/7/2021
• The Inria London Programme : kickoff, 1/2/2021
• « La révolution de l’apprentissage profond »,(in French) Yoshua Bengio, Interstices, 25/1/2019