Files

52: Machine Learning at the UTC-Heudiasyc lab

Professor Philippe Bonnifait, also Vice-Chairman of the scientific council of the University of Technology of Compiègne (UTC) former Director of a CNRS research group (GDR) in robotics between 2013 and 2017. Since January 2018, he has been Director of the UTC-Heudiasyc Laboratory, created in 1981. This state-of-the-art laboratory houses, in particular, the CID (Knowledge, Uncertainties, Data) team dedicated to research in artificial intelligence.

52: Machine Learning at the UTC-Heudiasyc lab

Modelling uncertainties in intelligent systems

Professor Thierry Denoeux has been a senior member of the Institut universitaire de France (IUF) since October 2019. Director of the Labex Maîtrise des Systèmes de Systèmes Technologiques (MS2T) within Heudiasyc, a joint UTC/CNRS unit, a researcher in the Knowledge, Uncertainties, Data (CID) team, he is also Editor-in-chief of two international journals: International Journal of Approximate Reasoning on uncertain reasoning and Array, a new open-access journal covering the entire field of computer science, launched in October 2018.

As a graduate engineer from École des Ponts ParisTech (ENSP), Thierry Denoeux pursued a PhD thesis on "the reliability of rain forecasts by weather radar" in a laboratory dedicated to the environment at ENPC. He is interested, among other things, in computer science, pattern recognition and image processing. "The goal was to process radar images to analyse and extrapolate the movement of rain cells for quantitative forecasting in the very short term (one to two hours). These predictions were used to optimize the management of large sewer systems in order to limit flooding in the event of a storm," he explains.

An interest that naturally led him, after his PhD, to join the Laboratoire d'Informatique Avancée de Compiègne (LIAC), Lyonnaise des Eaux, which has since become Suez, as a research engineer. He stayed there for three years and worked on European projects with lecturer cum-research scientists working at UTC, in the early 1990s, when artificial intelligence (AI) was already arousing a lot of interest with the development of expert systems.

He joined UTC Compiegne in 1992 as a contract lecturer-cum-research singulier scientist at Heudiasyc before being appointed as full professor in 1999. Several responsibilities followed suit: director of a joint laboratory with Suez, deputy director of Heudiasyc, vice-president of the scientific council of the UTC, scientific coordinator, before taking over as director in January 2019, of the Labex Maîtrise des Systèmes de Systèmes Technologiques (MS2T) - a tenyear project - which, as part of the Government Incentive Programme Investments for the Future (PIA), running until 2021. Also in January 2019, he took over the management of the SHIC¹ research federation, a CNRS structure initially grouping together the mixed Heudiasyc, BMBI and Roberval units, which were recently joined by the Costech unit. This federation provided impetus to a new dynamic for interdisciplinary technological research within the UTC.

At UTC-Heudiasyc, Thierry Denoeux is part of the CID team in charge of artificial intelligence, structured around two main research areas. The first concerns knowledge and data processing with themes such as knowledge modelling, machine learning and uncertainty management, a major challenge in both artificial intelligence and statistics. "Indeed, how can we model uncertainty, reason and make decisions knowing that we do not dispose of all the information needed ?" he says. The second area of research is concerned with customized adaptive systems. In other words, everything that relates to the interaction between humans and systems with the idea of designing systems that can automatically and dynamically adapt to the user and the context of use.

Thierry Denoeux works mainly on the first priority theme. "I work essentially on the modelling of uncertainties in intelligent systems, a theme that lies at the interface between artificial intelligence and statistics. I am particularly interested in the theory of belief functions, a theory of uncertainty that allows us to reason and make decisions in the presence of uncertainties. It is a general theory, which encompasses probability theory, and has many applications because uncertainties are ubiquitous. Research in this area is multidisciplinary and involves economists, AI specialists, statisticians and others," he explains.

A field that led him, in 2010, to participate in the creation of a learned society Belief functions and Applications society (BFAS), an association of which he is the President. The objective? In particular, to promote teaching, research, the advancement of knowledge in the field of belief functions and to explore the links with other theories of uncertainty. Hence the launch of international conferences held every two years - the next one will be convened in Shanghai in 2020 - and a thematic school for the training of PhD students, the latest edition of which was held in October 2019 in Siena (Italy).

However, Thierry Denoeux does not confine himself to the theoretical aspect of his research on belief functions, as he is also interested in the concrete applications that can derive from them. One example is automated postal address recognition, which was the subject of a CIFRE thesis in partnership with Solystic, one of the world leaders in the provision of automated sorting and distribution preparation solutions for parcels and mail. "This company sells machines with handwritten address recognition software. So when the address is not recognized, the envelope is rejected and processed manually. The challenge is to reject as few envelopes as possible while making as few errors as possible on those that are accepted. To meet these two criteria and improve machine performance, the idea was to integrate several software programs and combine the results of these systems using belief function theory," he explains.

Other applications include the work carried out with the French Institute of Transport, Planning and Network Science and Technology (IFSTTAR) and the SNCF on "Diagnosis of railway track circuits", and the ongoing collaboration with the Laboratory of Computer Science, Information Processing and Systems (LITIS) of the University of Rouen on "Segmentation of tumours in medical images and prognosis based on the evolution of patient data".

The theoretical corpus of belief functions is of course also of interest to the SyRI (Robotic Systems in Interaction) team, which is working in particular on intelligent vehicles (IV). "One of the problems in IV concerns perception. IVs are full of sensors and the challenge is to be able to process the information collected by these sensors to recognize objects on the road such as pedestrians, cyclists, etc. One of the problems in IVs is perception. We therefore need to combine the information from these different sensors. And here again, there is a lot of uncertainty, because each sensor provides partial and sometimes unreliable information about the environment," adds Thierry Denoeux.

Should we be afraid of AI? "Irrational fear is irrelevant. However, some applications of AI pose ethical problems, such as generalised video surveillance with, in particular, the development of facial recognition. As biologists have been doing for a long time, computer scientists must now be concerned about the ethical implications of their work," he concludes.