Julien Mairal, an innovator in machine learning

Julien Mairal has the sort of service record that many researchers would dream of having at the end of their careers. He has co-authored more than 100 papers, some of which have gone on to be internationally acclaimed, and has twice been awarded an ERC grant, a highly selective stipend given out by the European Research Council to leading scientists. At the Grenoble Alpes University Inria centre he is in charge of Thoth, a project team made up of around thirty or so individuals.

The Inria - French Academy of Sciences Young Researcher Prize is recognition for everything Julien Mairal has accomplished so far, with so much of his career still in front of him. “This is a prestigious award, with a lot of visibility in France. It’s in recognition of the diversity of my research: I never wanted to confine myself to any one subject.”

Pushing the boundaries of machine learning

Returning to France in 2012, the young researcher went back to Inria, working as a research fellow in Grenoble.

I knew that this would be a positive environment providing proper support for my research.

Julien Mairal joined Lear, a project team specialising in teaching machines visual models using large quantities of data. It wasn’t long before he was collaborating with other disciplines. A good example of this is the work he did with researchers in genomics at the Curie Institute and CNRS Lyon, helping them to put RNA strands back together following a sequencing operation that had chopped them up into tiny pieces.

In 2015 he and his colleagues published a first version of Catalyst, an algorithm which boosts the performance levels of certain machine learning tasks, which he would continue to fine-tune up until 2019. “That was a major breakthrough in my career: Catalyst solved a previously open problem regarding the acceleration of various algorithms, which had a significant impact.” 

The following year Julien Mairal was awarded an ERC grant by the European Research Council, enabling him to recruit a number of researchers to work alongside him. Two years later he became head of his own project team, which was now called Thoth.

Not long after that, he worked with a postdoctoral researcher he was supervising to develop the algorithm SwAV, followed by DINO in 2021, as part of a collaboration with Facebook. Both of these algorithms relate to what is known as self-supervised learning. “This involves teaching a system to automatically extract the information from scenes that is most relevant for visual recognition tasks, without the need for manual annotation.”   

This process is much quicker and more cost-effective than the traditional method, which involves giving a system millions of annotated images indicating whether an image is a chair, a tree, a child, etc.

The following year Julien Mairal was awarded an ERC grant by the European Research Council, enabling him to recruit a number of researchers to work alongside him. Two years later he became head of his own project team, which was now called Thoth.

Not long after that, he worked with a postdoctoral researcher he was supervising to develop the algorithm SwAV, followed by DINO in 2021, as part of a collaboration with Facebook. Both of these algorithms relate to what is known as self-supervised learning.

This involves teaching a system to automatically extract the information from scenes that is most relevant for visual recognition tasks, without the need for manual annotation.

This process is much quicker and more cost-effective than the traditional method, which involves giving a system millions of annotated images indicating whether an image is a chair, a tree, a child, etc.

Continuing his quest for multidisciplinarity, around this time Julien Mairal also began collaborating with astrophysicists from the CNRS based in Lyon and Paris. The aim was to improve the detection of exoplanets through the use of machine learning algorithms on images from large telescopes.

Today Julien Mairal’s diary is just as busy. He was recently awarded a second ERC grant, the goal this time being to encode physical principles into machine learning algorithms in order to simplify the learning phase.

He also provides scientific support to Enhance Lab, a startup launched by one of his PhD students in 2022. “It uses software to process streams of raw smartphone images in order to generate a high-definition view.” Yet another hurdle for this passionate researcher to overcome.

• Profiles of ERC award winners in France: Julien Mairal, IT researcher at INRIA, Horizon Europe, 23/3/2023.
• Julien Mairal and Liva Ralaivola - Artificial intelligence in all its forms (podcast), Institut Henri Poincaré, November 2021.
• Julien Mairal: developing “simplicity” in machine learning, Inria, 31/1/2023.
• Transforming data into knowledge: Julien Mairal's ERC project, Inria, 6/10/2016.