Eight IEA projets selected at CNRS Informatics

One of important tools for compressing and dealing with high-dimensional data are tensor decompositions, which  express multidimensional arrays (tensors) using compact low-rank representations.

Low-rank and tensor techniques are successfully employed in a wide range of applications. They are particularly relevant for methods of AI (artificial intelligence) operating with multidimensional arrays of data and model weights, in order to reduce computational and memory requirements of such methods (frugal AI).

This project explores advanced (coupled, multi-layer, etc.) tensor decompositions and approximations appearing in the context of several applications, such as neural network compression, anomaly detection in hyperspectral imaging, and analysis of functional magnetic resonance imaging (fMRI) data. The main objective is to conceive algorithms and study the key property of identifiability for such models, which is crucial for applications. The project is led by Konstantin Usevich, CNRS researcher at Centre de recherche en automatique de Nancy (CRAN – CNRS/Université de Lorraine) and aims at developing a collaboration of the team SiMul of CRAN with experts on this topic from the Department of Computer Science of KU Leuven (Belgium).

Artificial intelligence (AI) is pervasive in biology. From clinical decision support to the modeling of biological mechanisms, these systems can generate complex predictions that often elude human intuition. As a result, they are frequently described as "black boxes" due to the difficulty of understanding their internal logic. This opacity often leads to confusion, making AI interpretability a priority to address the security, ethical, and scientific challenges posed by these methods.

"Predicting is not explaining" said René Thom. This is precisely the goal of DecipHear: to develop interpretability tools capable of deciphering the nonlinear strategies of neural networks simulating auditory brain activity. At the intersection of computer science and neuroscience, the project is led by Étienne Thoret, CNRS researcher at Institut de neurosciences de la Timone (INT - CNRS/Aix-Marseille University) in collaboration with Frédéric Gosselin, professor at the University of Montreal.

The idea? To use two methods from cognitive neuroscience to probe AI systems, aiming to identify the key information driving predictions and to shed light on the sound representations encoded both in artificial networks and in the human brain. The anticipated breakthroughs will also support the development of new explainability methods, applicable to other AI-related challenges.

Formal program verification uses logic to demonstrate mathematically that programs satisfy certain properties. This project focuses on the verification of probabilistic programs, i.e. those using probabilistic primitives such as random draws. This programming paradigm is widely used in fields such as cryptography and privacy protection. However, in the deterministic (non-probabilistic) context, two classical approaches to formal verification of deterministic programs that can be automated are algebraic approaches, for programs written in imperative languages, and type systems, for functional languages. The project will explore extensions of these two categories of methods to probabilistic programs, as well as their applications to differential privacy, a privacy protection method based on perturbing the computation by adding probabilistic noise.

This project, led by Patrick Baillot, CNRS senior researcher at the Centre de recherche en informatique, signal et automatique de Lille (CRIStAL - CNRS/Centrale Lille/Université de Lille), will draw on the complementary skills in logic and semantics of programming languages of participants from the CRIStAL laboratory, and in verification and differential confidentiality of participants from Boston University.

Modern computer networks increasingly rely on machine learning (ML) techniques for a wide range of management tasks, from security to performance optimization. While labeled network traces are essential for these tasks, a significant challenge in training network-focused ML models is the scarcity of labeled network datasets. Although synthetic network traces can augment existing datasets, current techniques typically produce only aggregated flow statistics or limited packet attributes. These approaches fall short when model training requires features that are only available from complete packet traces. This limitation manifests in both insufficient statistical fidelity to real traces and suboptimal performance on ML tasks when used for data augmentation.

This project explores the application of generative AI to address emerging challenges in computer systems, particularly in computer networks. Specifically, the project will investigate the development of generative artificial intelligence models to create high-resolution synthetic network traffic traces. The research will be conducted by a French-US collaboration between Francesco Bronzino, associate professor at ENS de Lyon and member of the Laboratoire de l'informatique du parallélisme (LIP – CNRS/ENS de Lyon/Université Claude Bernard), and Nick Feamster, Neubauer professor at the University of Chicago.

Graphs (or networks) are popular mathematical models for representing interrelated data of any kind. The homomorphism of a graph G usually helps embed it in another graph H (smaller or bigger than G) while preserving certain important properties. The term Col(G,H) denotes the space of all homomorphisms of G to H, and interestingly, Col(G,H) is also viewed as a graph. 

HoDReCo is a project led by Moritz Mühlenthaler, associate professor at Institut Polytechnique de Grenoble and member of the laboratory Sciences pour la conception, l'optimisation et la production de Grenoble (G-SCOP – CNRS/Université de Grenoble). His primary objective is to study different structural and algorithmic aspects of Col(G,H), with a particular focus on the emerging topic of reconfiguration and the classical topic of coloring. These studies of structural and algorithmic aspects of Col(G,H) finds its applications in CSP (Constraint Satisfaction Problem), statistical physics, combinatorial games, scheduling, wireless networks, and graph database, to name a few. His study is expected to impact both the theory building and problem solving sides of the focus areas.

Motion cOmprehensive predicTIon Framework 

Predicting human gestures is essential to making assistive robots safe and efficient. Inverse optimal control aims to identify, from observed movements, the cost functions that humans seek to minimize (energy, comfort, etc.).

The research project is led by Vincent Bonnet from the Laboratoire d'analyse et d'architecture des systèmes (LAAS-CNRS), associate professor at the University of Toulouse and member of the International research laboratory IPAL, Dana Kulic, professor of robotics and director of the robotics centre at Monash University in Melbourne (Australia), and two PhD students. The main objective of this collaboration is to compare the methods of each team and then develop new inverse optimization algorithms to identify biomechanical cost functions from human gestures, in order to feed a knowledge model, or foundation model of human movement, capable of generating a large number of optimal trajectories close to those observed in humans. In contrast to the literature based on LLMs (Large Language Models), this approach maintains a degree of explicability in the prediction of human movement. This model can then be used to simulate or predict human movements in a variety of contexts, with the ultimate aim of adapting robot behavior to that of humans in a natural, anticipatory way.

The Laboratoire d'analyse et d'architecture des systèmes (LAAS-CNRS/INSA Toulouse/INP de Toulouse/Université de Toulouse) and the University of Seville are teaming up to explore new frontiers in control engineering. Through this collaboration, researchers from both sides will combine their strengths in hybrid dynamical systems — a powerful way to model complex behaviors that mix smooth changes and sudden jumps. This shared expertise opens up possibilities for more reliable power grids, efficient satellite maneuvers, and smarter energy systems — exactly the kind of resilience needed to prevent disruptions like the blackout in Spain in spring 2025.

PhD students and young researchers will also be deeply involved, gaining hands-on experience in cutting-edge projects and learning from experts in both France and Spain. This project led by Luca Zaccarian, CNRS senior researcher at LAAS-CNRS, is a true exchange of knowledge, with the goal of building not only better technologies but also the next generation of engineers ready to tackle big challenges.

Digital twins are experiencing unprecedented growth as part of Industry 4.0. TWIN4NET aims to transpose this innovative approach to wireless networks, such as 5G, Wi-Fi and the Internet of Things. A true optimization lever, the digital twin enables network infrastructure to be continuously adapted to changing conditions, while making it more economical in terms of energy and radio resources. By adjusting resources as closely as possible to actual needs, it becomes possible to limit infrastructure oversizing. The digital twin enables different configurations to be simulated and tested virtually on models, thus avoiding any disruption to network operations.

This project, led by Fabrice Theoleyre, CNRS research director at Laboratoire des sciences de l'ingénieur, de l'informatique et de l'imagerie (ICube - CNRS/Université de Strasbourg), will consolidate the collaboration between the two partners. The ICube laboratory's Networks team will contribute its expertise in wireless network design and experimentation (5G and Internet of Things), while the National University of Pusan (South Korea) will contribute its recognized know-how in Artificial Intelligence and optimization applied to networks.