DSAIDIS Chair: Data Science and AI for Industry

The rise of Big Data and artificial intelligence (AI) has opened up new prospects for companies. They can rapidly analyze large volumes of data, support real-time decision-making and make reliable forecasts. These recent advances offer unprecedented opportunities for many professional sectors.

However, among these, industry and services have certain specificities to which systems must adapt. In these sectors, the analysis of data collected (in considerable quantities) presents a number of challenges:

- contaminated data, noise, extreme or missing values

- conditions that change over time

- unlabeled or partially labeled data, i.e. with little associated information

- heterogeneity due, among other things, to the diversity of the sources used

- A sometimes insufficient quantity of data, forcing the system to generate new data or to operate with a low quantity.

Moreover, these obstacles arise in a context where, for industry in particular, systems need to prove that they are "trustworthy" if they are to be fully adopted. Systems must therefore meet high standards demonstrating their reliability, robustness and ability to be explained.

AI tools for real-life situations

To meet these challenges, the DSAIDIS Chair (Data Science & Artificial Intelligence for Digitalized Industry & Services) was created with the support of the TSN Carnot Institute. It brings together some twenty professors and researchers from Télécom Paris, who collaborate with the five industrial partners: Airbus Defence and Space, Engie, IDEMIA, Safran and Valeo. " Our links with the companies are strong, and each feeds off the other," says Pavlo Mozharovskyi, a teacher-researcher at Télécom Paris and co-leader of the Chair. " For our part, we draw inspiration for our work from their real-life problems. And when we respond to them, we propose systems capable of bringing them concrete added value. "

The ambition of the DSAIDIS Chair is to develop methodological tools that can be applied under realistic conditions. For each project, this approach requires a theoretical modeling phase, followed by the development of algorithms that make practical use of the results previously obtained. To this end, the researchers rely on machine learning methods, a branch of artificial intelligence.

" The adaptation of our systems to realistic conditions of use is at the heart of our approach," emphasizes Florence d'Alché, professor-researcher at Télécom Paris and holder of the Chair. " We don't apply our models to ideal cases, far removed from reality. On the contrary, they are designed to adapt to noisy environments, contaminated data, extreme values and so on. And all this while satisfying high reliability criteria and offering maximum theoretical guarantees. "

What's more, the tools developed are in line with a general current issue: sobriety. " We make sure that our models and algorithms require as little memory, computing power and data as possible ", asserts the researcher. This is still a recent concern in the AI community.

4 axes for greater confidence in AI

In detail, research is organized around four axes, defined in collaboration with the five industrial partners.

1) Time series analysis and forecasting

While this first line of research may seem classic, the researchers intend to approach it from a new angle, notably by combining classical statistical methods with machine learning tools. " The originality of our study also lies in the fact that we are not interested in a series of measurements at an instant T, but in considering a portion of a signal over an interval of time," explains Florence d'Alché. " An approach that offers additional tools for identifying properties relevant to decision-making. "

2) Large-scale exploitation of partially labeled and heterogeneous data

This area, which covers a broad scope, aims to address the issues associated with Big Data in industry and services. In particular, how can we effectively exploit large volumes of data, despite incomplete labeling and a wide diversity of sources?

3) Machine learning for reliable and robust decision-making

Here, the aim is to boost user confidence in AI tools. How can machine learning algorithms take account of data imperfections (noise, contamination, extreme values...), while demonstrating reliability? How can they help correct biases and ensure greater fairness? These are issues that go beyond technical problems, and involve researchers in the economic and social sciences.

Explainability is also a priority. " Current machine learning models are complex and operate like a 'black box'," notes Pavlo Mozharovskyi. " So we're looking to provide explanations for the decisions made by the machine, so that the 'black box' becomes, perhaps not white, but grey. " A wish strongly expressed by the Chair's industrial partners, and by society in general.

4) Learning in interaction with the environment

In practice, artificial intelligence systems have to integrate into changing environments. They must therefore be able to take these changes into account and adapt their operation accordingly. The research team is thus working to equip tools with an autonomous and continuous learning capability.

Fruitful collaboration between industry and academia

Several projects have already been carried out in collaboration with the Chair's industrial partners. Valeo, for example, called on the researchers to improve one of its production lines. Like all manufacturers, the company is constantly seeking to reduce its rate of defective parts, in order to optimize its output. " We began with a broad understanding of the data and processes involved, before proceeding to analyze them," explains the co-host of the DSAIDIS Chair. " Then, thanks to extensive data visualization work, we were able to identify the decisive parameters for avoiding defects during manufacturing. We then used statistical tools to draw up recommendations. These were then applied to the production line, with immediate results: the rate of defective parts was almost halved!

This success has led to the launch of a CIFRE thesis, still in progress, with the same partner, on another production line. Here, the aim is to explain anomalies occurring during manufacturing and identify their most likely sources. " Our statistical analysis enables us to determine which workstations on the production line are most likely to lead to part defects ", Pavlo Mozharovskyi sums up. " This enables the industrialist to know where to concentrate his efforts to improve his process. " A tool that Valeo's teams have already begun to use, in the field.

Researchers have also been working with IDEMIA, which develops biometric solutions, notably facial recognition systems. A technology that has often been criticized for its lack of fairness. " Our aim is precisely to correct selection bias," says the chairholder. " In concrete terms, our system aims either to reweight the training data, in order to blur the effects of under-representation of certain populations, or to ignore certain attributes (gender, skin color...). "

Through these examples, the DSAIDIS Chair is a perfect illustration of the synergy possible between the worlds of industry and academia. And despite an official closing scheduled for the end of 2023, the adventure is not set to end this year. " We're currently working on renewing the chair," says Florence d'Alché. " We're meeting with each of the partners to determine which of the subjects we've already studied deserve to be extended, to identify new issues, and to propose new areas of study. This process of reflection is also fuelled by meetings with new industrial partners, who could join the Chair as early as 2024.