TAU

TAU - 2025

2025Activity reportProject-TeamTAU‌

RNSR: 201622258D

Research center‌ Inria Saclay Centre at‌‌ Université Paris-Saclay
In partnership with:CNRS, Université Paris-Saclay‌
Team name: TAckling the‌ Underspecified
In collaboration with:‌‌Laboratoire Interdisciplinaire des Sciences du Numérique

Creation of‌ the Project-Team: 2019 July‌ 01

Each year, Inria‌‌ research teams publish an Activity Report presenting their‌ work and results over‌ the reporting period. These‌‌ reports follow a common structure, with some optional‌ sections depending on the‌ specific team. They typically‌‌ begin by outlining the overall objectives and research‌ programme, including the main‌ research themes, goals, and‌‌ methodological approaches. They also describe the application domains‌ targeted by the team,‌ highlighting the scientific or‌‌ societal contexts in which their work is situated.‌

The reports then present‌ the highlights of the‌‌ year, covering major scientific‌ achievements, software developments, or teaching contributions. When relevant,‌ they include sections on software, platforms, and open‌ data, detailing the tools developed and how they‌ are shared. A substantial part is dedicated to‌ new results, where scientific contributions are described in‌ detail, often with subsections specifying participants and associated‌ keywords.

Finally, the Activity Report addresses funding, contracts,‌ partnerships, and collaborations at various levels, from industrial‌ agreements to international cooperations. It also covers dissemination‌ and teaching activities, such as participation in scientific‌ events, outreach, and supervision. The document concludes with‌ a presentation of scientific production, including major publications‌ and those produced during the year.

Keywords

Computer‌ Science and Digital Science

A3.3.3. Big data analysis‌
A3.4. Machine learning and statistics
A3.5.2. Recommendation systems‌
A6.2. Scientific computing, Numerical Analysis & Optimization
A8.2.‌ Optimization
A8.6. Information theory
A8.12. Optimal transport
A9.2.‌ Machine learning
A9.3. Signal processing

1 Team‌ members, visitors, external collaborators

Research Scientists

Cyril Furtlehner‌ [Team leader, INRIA, Researcher,‌ Interim, HDR]
Guillaume Charpiat [INRIA‌, Researcher]
Thi Tuyet Trang Chau [‌INRIA, Advanced Research Position, from Feb‌ 2025]
Alex Davey [INRIA, Starting‌ Research Position, from Oct 2025]
Flora‌ Jay [CNRS, Researcher]
Matthieu Nastorg‌ [Augur, Industrial member, from Sep‌ 2025]
Stephane Rivaud [INRIA, Starting‌ Research Position]
Marc Schoenauer [INRIA,‌ Senior Researcher, Emeritus, HDR]
Martine‌ Michele Sebag [CNRS, Senior Researcher,‌ Emeritus, HDR]
Alena Shilova [INRIA‌, ISFP]

Faculty Members

Philippe Caillou [‌UNIV PARIS SACLAY, Associate Professor]
Sylvain‌ Chevallier [UNIV PARIS SACLAY, Professor]‌
Sergio Chibbaro [UNIV PARIS SACLAY, Professor‌, HDR]
Cécile Germain [UNIV PARIS‌ SACLAY, Emeritus, HDR]
Matthieu Kowalski‌ [UNIV PARIS SACLAY, Associate Professor Delegation‌]
François Landes [UNIV PARIS SACLAY,‌ Associate Professor]

Post-Doctoral Fellows

Hugo Niels Corentin‌ Brehier [CENTRALESUPELEC, until Nov 2025]‌
Thi Tuyet Trang Chau [INRIA, from‌ Feb 2025]
Shuyu Dong [CENTRALESUPELEC,‌ until Jun 2025]
Matthieu Nastorg [INRIA‌, Post-Doctoral Fellow, until Jun 2025]‌
Lisheng Sun [UNIV PARIS SACLAY, Post-Doctoral‌ Fellow, from Feb 2025 until Aug 2025‌]
Naama Wagner [UNIV PARIS SACLAY,‌ Post-Doctoral Fellow, from Apr 2025]

PhD‌ Students

Anaclara Alvez Canepa [UNIV PARIS SACLAY‌]
Bruno Aristimunha Pinto [INRIA]
Nicolas‌ Bereux [UNIV PARIS SACLAY]
Eva Boguslawski‌ [RTE, CIFRE]
Styliani Douka [‌INRIA]
Romain Egele [Argone National Lab.‌, until Sep 2025]
Theofanis Ifaistos [‌INRIA, from Feb 2025]
Thanh Gia Hieu Khuong [UNIV‌ PARIS SACLAY, from‌ Apr 2025]
Ismail‌‌ Labiad [Meta, CIFRE, from Mar‌ 2025]
Alice Athenais‌ Lacan [Université Paris-Saclay‌‌]
Jean-Baptiste Malagnoux [INRIA, from Oct‌ 2025]
Jazeps Medina‌ Tretmanis [UNIV BROWN‌‌]
Lorenzo Mensi [UNIV PARIS SACLAY,‌ from Oct 2025]‌
Solal Nathan [UNIV‌‌ PARIS SACLAY]
Alfonso De Jesus Navas Gomez‌ [INRIA, from‌ Jun 2025 until Aug‌‌ 2025, Invited]
Audrey Poinsot [EKIMETRICS‌, CIFRE]
Arnaud‌ Quelin [SORBONNE UNIVERSITE‌‌, until Sep 2025]
Cyriaque Rousselot [‌UNIV PARIS SACLAY]‌
Theo Rudkiewicz [ENS‌‌ PARIS-SACLAY]
Nilo Schwencke [Université Paris-Saclay]‌
Dylan Sechet [UNIV‌ PARIS SACLAY, from‌‌ Aug 2025]
Antoine Szatkownik [UNIV PARIS‌ SACLAY]
Luca Teodorescu‌ [UNIV PARIS SACLAY‌‌, from Sep 2025]
Sébastien Velut [‌UNIV PARIS SACLAY]‌
Manon Verbockhaven [INRIA‌‌, until Mar 2025]
Thomas Weikert [‌INRIA]
Assia Wirth‌ [UNIV PARIS SACLAY‌‌, until Jan 2025]

Technical Staff

Anne-Catherine‌ Letournel [CNRS,‌ Engineer]
Dylan Sechet‌‌ [CENTRALESUPELEC, Engineer, until Aug 2025‌]

Interns and Apprentices‌

Valerio Actis Dato Casale‌‌ [INRIA, Intern, from Mar 2025‌ until Jul 2025]‌
Ilan Aliouchouche [UNIV‌‌ PARIS SACLAY, until Apr 2025]
Mathias‌ Bichon [INRIA,‌ Intern, from Apr‌‌ 2025 until May 2025]
Leo Burgund [‌INRIA, Intern,‌ from Apr 2025 until‌‌ Aug 2025]
Luca Camagna [INRIA,‌ Intern, from Mar‌ 2025 until Jul 2025‌‌]
Jan Chelmecki [UNIV PARIS SACLAY,‌ Intern, from Jun‌ 2025 until Jul 2025‌‌]
Awa Cisse [Institute for Field Education‌, Intern, from‌ Jun 2025 until Aug‌‌ 2025]
Felix Houdouin [INRIA, Intern‌, from May 2025‌ until Sep 2025]‌‌
Joachim Jobard [INRIA, Intern, from‌ Oct 2025]
Julien‌ Marie-Anne [INRIA,‌‌ Intern, from Apr 2025 until Oct 2025‌]
Lorenzo Mensi [‌Inria, Intern,‌‌ from Feb 2025 until Jun 2025]
Andrei-Tiberiu‌ Pantea [Unviersité Paris-Saclay‌]
Marina Pereira–Garcia [‌‌Université McGill]
Eliott Pradeleix [LISN,‌ Intern, from Apr‌ 2025 until Jul 2025‌‌]
Luca Teodorescu [INRIA, Intern,‌ from Apr 2025 until‌ Sep 2025]
Jeremie‌‌ Touati [INRIA, Intern, from Apr‌ 2025 until Aug 2025‌]
Sheng Wan [‌‌ECOLE POLY PALAISEAU, Intern, from Jul‌ 2025 until Aug 2025‌]
Assia Wirth [‌‌UNIV PARIS SACLAY, from Feb 2025]‌
Leonardo Zanini [UNIV‌ PARIS SACLAY, Intern‌‌, from Jun 2025 until Jul 2025]‌

Administrative Assistant

Julienne Moukalou‌ [INRIA]

Visiting‌‌ Scientist

Julien Mille [CNRS, from Sep‌ 2025]

External Collaborators‌

Aurelien Decelle [Universidad‌‌ Polit´ecnica de Madrid,‌ Associate member]
Alessandro Ferreira Leite [INSA‌ ROUEN NORMANDIE]
Isabelle Guyon [UNIV PARIS‌ SACLAY]
Lionel Mathelin [CNRS, HDR‌]
Thibault Monsel [Univ Paris Saclay,‌ until Oct 2025]
Leo Benoit Planche [‌UNIV PARIS SACLAY, from Jun 2025]‌
Onofrio Semeraro [CNRS]
Beatriz Seoane Bartolomé‌ [Universidad Complutense de Madrid]
Burak Yelmen‌ [UNIV PARIS SACLAY]

2 Overall objectives‌

2.1 Presentation

Building upon the expertise in machine‌ learning (ML) and stochastic optimization, and statistical physics‌ of the former TAO project-team, the TAU team‌ aims to tackle the vagueness of the Big‌ Data purposes. Based on the claim that‌ (sufficiently) big data can to some extent compensate‌ for the lack of knowledge, Big Data is‌ hoped to fulfill all Artificial Intelligence commitments.

This‌ makes Big Data under-specified in three respects:

A‌ first source of under-specification is related to common‌ sense, and the gap between observation and‌ interpretation. The acquired data do not report on‌ “obvious” issues; still, obvious issues are not necessarily‌ so for the computer. Providing the machine with‌ common sense is a many-faceted, AI hard, challenge.‌ A current challenge is to interpret the data‌ and cope with its blind zones (e.g., missing‌ values, contradictory examples, ...).
A second source of‌ under-specification regards the steering of a Big Data‌ system. Such systems commonly require lifelong learning‌ in order to deal with open environments and‌ users with diverse profiles, expertises and expectations. A‌ Big Data system thus is a dynamic process,‌ whose behavior will depend in a cumulative way‌ upon its future environment. The challenge regards the‌ control of a lifelong learning system.
A‌ third source of under-specification regards its social acceptability.‌ There is little doubt that Big Data can‌ pave the way for Big Brother, and ruin‌ the social contract through modeling benefits and costs‌ at the individual level. What are the fair‌ trade-offs between safety, freedom and efficiency ? We‌ do not know the answers. A first practical‌ and scientific challenge is to first assess, and‌ then enforce, the trustworthiness of solutions.

However,‌ several concerns have emerged in the last years‌ regarding Big Data models. First, in industrial context,‌ data is now always big, and many practical‌ problems are relevant to small data. On‌ the opposite, when big data is available, the‌ arms race around LLMs has given birth to‌ increasingly big models, involving hundreds of billions of‌ parameters, and environmental concerns are becoming increasingly high,‌ for their training, but even for their use‌ and the inference process.

Our initial overall under-specification‌ considerations, mitigated with the concerns above, have lead‌ the team to align its research agenda along‌ four pillars:

Frugal Learning, addressing the environmental concerns,‌ in terms of deep network architecture and considering‌ the small data regimes;
Causal Learning, a grounded‌ way to address the trustworthiness issue by improving explainability of the results;‌
Bidirectional links with Statistical‌ Physics, to better understand‌‌ very large systems and improve their performances, both‌ in terms of accuracy‌ of the models and‌‌ energy consumption in their use;
Hybridization of Machine‌ Learning with Numerical Simulations,‌ again aiming to reach‌‌ better efficiency while decreasing the computing needs.

Last‌ but not least, the‌ organization of challenges and‌‌ the design of benchmarks, a cornerstone of Machine‌ Learning nowadays, remains an‌ active thread of the‌‌ team activity, in particular through the Codalab platform‌ and its new version‌ Codabench.

3 Research‌‌ program

3.1 Frugal Learning

Frugality is a must‌ for machine learning: because‌ of scientific concerns (monster‌‌ models imply non-reproducible science); because of sustainability concerns‌ (energy consumption to train‌ and use models); because‌‌ of applicability concerns: in most non-GAFAM/GAMAM settings, we‌ deal with small data,‌ and PhD students not‌‌ infrequently receive the promised data in the last‌ months of their PhDs.‌

We target in particular‌‌ three domains: data frugality, computational complexity at test‌ time (to minimize environmental‌ footprint when using the‌‌ trained network at large scales), and computational complexity‌ of neural architecture search‌ (i.e. of the automatically‌‌ finding of neural architectures suitable for a given‌ machine learning task at‌ hand, at training time).‌‌ The mainstream strategy suggests finding a model in‌ a large (overparameterized) model‌ space, in order to‌‌ avoid optimization and expressivity issues, and then pruning‌ it 109. An‌ alternative to the above‌‌ strategy, named neural network growth, consists in starting‌ from a tiny architecture‌ and grafting additional neurons‌‌ or layers to extend its representation power on‌ demand, on the fly‌ during training. This raises‌‌ interesting mathematical questions regarding optimization, generalization, and statistical‌ significance.

An approach we‌ are currently developping follows‌‌ the preliminary proof of concept in M. Verbockhaven's‌ PhD where we seek‌ to adapt the neural‌‌ tangent kernel to the directions desired by the‌ functional gradient descent. This‌ kind of approach could‌‌ be useful not only to automatically (and frugally)‌ design from scratch a‌ neural network architecture suitable‌‌ for a new task, but could also be‌ of prime interest in‌ classical Neural Architecture Search‌‌ to provide directly optimal architecture variations instead of‌ searching for them in‌ a computationally-heavy trial-and-error fashion.‌‌

A nice byproduct is that by building smaller‌ models, one potentially requires‌ smaller data, and is‌‌ potentially less prone to overfit. This opens interesting‌ questions regarding regularization in‌ deep learning and advocates‌‌ for a more reasonable, guided use of combinatorics,‌ that appear through traditional‌ random initializations of numerous‌‌ neurons (lottery ticket hypothesis 94).

3.2 Causal‌ Learning

The rise of‌ causal modelling (CM) has‌‌ an impact on the general agenda of machine‌ learning, more aware of,‌ and more robust w.r.t.‌‌ the potential and usual differences of data distributions‌ between training and testing‌ times or along lifelong‌‌ learning. This new agenda focuses on sorting out‌ distribution-independent relations (hopefully causal‌ ones) among the observed‌‌ features, and other relations,‌ possibly reflecting spurious correlations. The expected benefits of‌ this causality-inspired focus is to deliver learned models‌ that are: i) more robust w.r.t. the non‌ iid setting; ii) more interpretable; iii) possibly humanly‌ verified. The last two properties only hold, naturally,‌ if the features are expressed at a sufficient‌ level of generality.

A key scientific question is‌ whether and how the main lesson of Deep‌ Learning (It's the representation, stupid !) can be‌ ported to causal modelling, particularly so when dealing‌ with raw, redundant and/or high dimensional data. The‌ use of latent variables and structures in e.g.‌ 97, 119, 121 has shown its‌ potential to disentangle root causes (sources of the‌ observed data) and cope with hidden confounders. However,‌ causal modelling comes with the key requirement of‌ identifiability/uniqueness of the learned causal models, that is‌ in general not satisfied in mainstream machine learning.‌

A promising research direction toward model identifiability is‌ to investigate the stability of causal discovery. Formally,‌ one might want that, if data $𝒟$ yields‌ model $𝒢$ , then data $𝒟^{'}$ generated‌ after $𝒢$ yields a model $𝒢^{'}$ that‌ is in essence same as $𝒢$ . This‌ direction opens to two strategies: i) observing the‌ differences between $𝒢$ and $𝒢^{'}$ sheds some‌ light about the diversity in the data with‌ some/no impact on the causal modelling output, i.e.‌ the biases of the causal discovery algorithms; ii)‌ and more deeply, the issue of stability can‌ inspire new learning criteria, enforcing the stability of‌ the causal models under such changes of distribution.‌ Another hot research direction investigates how to improve‌ the interpretability of a model, without degrading too‌ seriously its accuracy. Let us focus on the‌ task of interpreting hidden variables and their interactions.‌ A possible strategy $-$ at the core of‌ the AI2 French-German proposal, 2023-2026; coll. Fraunhofer Bonn‌ $-$ takes inspiration from the Multi-Criteria Decision Aid‌ literature (and the lessons learned in R. Bresson's‌ PhD 77, 78). The idea is‌ that i) if the last say two layers‌ of a deep net were structured as a‌ hierarchical choquet integral (HCI); ii) and if their‌ input (the nodes in the layer before) were‌ interpretable (giving a feature name to each node),‌ then the black box could be made transparent,‌ expressing sparse hierarchical interactions (HCI) of these features.‌ The first condition can be handled by retraining‌ a trained efficient deep net, and imposing HCI‌ constraints on the last two layers. A pending‌ question is how these constraints would degrade the‌ loss accuracy (depending on the number of would-be‌ features). The second condition will be met by‌ associating a supervised binary learning problem to each‌ node, and involving the expert in the loop‌ (or possibly exploiting textual information about the samples)‌ to solve it.

3.3 Machine Learning with/for Statistical‌ Physics

Concerning the links between statistical physics and‌ machine learning, we are working on both aspect of ML with Statistical‌ Physics and Statistical Physics‌ for ML.

1- The‌‌ first line of research, based on our expertise‌ on generative models, will‌ be headed toward efficient‌‌ methods for frugal and interpretable generative models, typically‌ energy based models (EBM)‌105. In particular‌‌ concerning explainability we will look for physically-inspired interpretable‌ feature extraction processes, exploring‌ the possibilities of using‌‌ EBMs as data-driven fitness landscapes.

2- This explainability‌ aspect will be actually‌ important for our second‌‌ axes concerning applications of EBMs in bioinformatics. For‌ instance, given data of‌ protein's families with common‌‌ ancestors, we expect to be able to learn‌ a model describing the‌ statistics of the family,‌‌ and then use this model to predict the‌ mutation of the amino-acid.‌ More broadly we will‌‌ develop methods for direct coupling extraction with RBMs,‌ clustering of data in‌ families and subfamilies, semi-supervised‌‌ strategies and use EBM for pattern extraction in‌ genomics/proteomics sequence datasets.

3-‌ Our third axes will‌‌ focus on symmetries both for methods and applications.‌ "It is only slightly‌ overstating the case to‌‌ say that physics is the study of symmetry"‌ (Philip Anderson 1972), and‌ enforcing symmetries into models‌‌ or finding symmetries in the data108 is‌ also key to ML.‌ CNNs can enforce translation‌‌ equivariance, GNNs enforce permutation equivariance, and more recently,‌ rules for building roto-translation-equivariant‌ networks have been devised‌‌82. The importance of symmetries has been‌ acknowledged in 79,‌ coining the term “geometric‌‌ deep learning” to refer to group-invariance aware neural‌ networks. We are working‌ on pushing roto-translation equivariance‌‌ further, with application to molecular systems or amorphous‌ materials. Furthermore, from statistical‌ physics we know that‌‌ sytems display scale-invariant distributions at their critical point.‌ Starting from simple avalanche‌ models as benchmarks, we‌‌ want to design networks that would be genuinely‌ scale-equivariant (or invariant). Applications‌ range from seismic hazard‌‌ to solar wheather forecast, i.e. any area where‌ large events-related data are‌ scarce. Such networks would‌‌ de facto perform extrapolation, a rare feature in‌ Machine Learning. This avenue‌ of research is being‌‌ studied within Anaclara Alvez' PhD (co-supervised by Cyril‌ Furtlehner and François Landes)‌ and the ANR Scalp‌‌ (2025-2028) to extend this to mult-fractal data.

4-‌ Our last axes deals‌ with fundamental properties of‌‌ ML like for instance neural scaling laws124‌ and is based on‌ recent theoretical progresses like‌‌ the formulation of the neural tangent kernel99‌ and the lazy regime‌81. Various asymptotic‌‌ results can be obtained thanks to random matrix‌ theory or replica approaches.‌ Equipped with such tools‌‌ we would like to explore for instance the‌ learning dynamics beyond the‌ lazy regime, the out-of-equilibrium‌‌ regimes of EBMs via dynamical mean field theory‌ but also the utility-privacy‌ trade-off with solvable models.‌‌

3.4 Machine Learning for Numerical Simulations

Until recently,‌ applying off-the-shelf neural nets‌ to numerical simulations (e.g.,‌‌ approximating the solution of PDEs) could only compete‌ with numerical solvers in‌ a few situations: when‌‌ the problem is simple‌ and of reasonable size, and when a limited‌ accuracy, that does not need to be guaranteed,‌ is sufficient. For instance, in cases involving chaotic‌ behaviors (e.g. turbulent flows in fluid dynamics), current‌ models fail to fit the target trajectory in‌ the mid to long term. The situation is‌ rapidly evolving (see e.g., GraphCast, by DeepMind 103‌), but there remains a need for tighter‌ coupling between ML and simulations.

Building upon TAU‌ expertise in numerical engineering, it is suggested that‌ the diversity of use cases tackled in applications‌ (recent and on-going PhDs of W. Liu, E.‌ Menier, M. Nastorg, E. Goutierre; T. Monsel, and‌ collaboration with the IRT SystemX IA2 program as‌ well as with IFPEN) can lead to formulating‌ general principles and methodology.

One research direction is‌ to consider more structured losses/architectures. This research direction‌ evolves at a rapid pace: from convolutional architectures,‌ to distributional architectures enforcing invariance or equivariance properties‌ 83, to optimal transport based embeddings 115‌. It is believed that new losses, aimed‌ at preserving statistical quantities (e.g. high order moments;‌ extreme value exponents), might help to learn and‌ reproduce chaotic data trajectories, better than MSE losses.‌ Nevertheless, until theoretical guidelines are available to the‌ practionner, it is important to be able to‌ experimentally guide and validate users' choices in terms‌ of architecture/loss, for any new use-case. There is‌ today a lack of well-grounded and widely accepted‌ benchmarks, and we contribute to the IRT SystemX‌ LIPS platform (Learning Industrial Physical Simulation benchmark suite)‌ 107, lead by our collaborators from IRT‌ (Mouadh Yagoubi) and RTE (Benjamin Donnot and Antoine‌ Marot).

Another direction of research concerns how the‌ domain know-how can best be conveyed to the‌ learning process: through priors; or warm-starting the solution;‌ or enforcing the required solution properties through specific‌ loss terms; or maybe simply choosing the right‌ training samples.

A theoretically and practically important domain‌ concerns the coupling of an ML model and‌ a numerical simulator, with mutual benefits (compensating for‌ insufficient data; adjusting the simulator hyper-parameters; prioritizing new‌ experiments toward optimal design or model identification; providing‌ a fast sampler; addressing inverse problems). Mimicking the‌ structure of the simulator/the physical phenomenon through the‌ neural architecture helps to guide the optimization, all‌ the more so as it supports the definition‌ of auxiliary losses (e.g. based on internal states‌ of the simulator). Again, the use of auxiliary‌ losses can be very useful, if an appropriate‌ learning schedule has been defined (controlling the impact/weight‌ of each auxiliary loss depending on the current‌ state of the model and of the learning‌ trajectory).

Last but not least, unleashing the power‌ of the recently emerged Foundation Models and Transformers‌ resulted in low hanging fruits (e.g., more powerful‌ surrogate models) that have not yet been picked‌ up, and will also open new avenues for‌ hybrid/multidisciplinary research.

3.5 Challenge Organization

In the rapidly‌ evolving field of machine learning (Data-Driven Artificial Intelligence) empirical evaluations of new‌ algorthms to confirm their‌ effectiveness and reliability is‌‌ even more essential. This trend is intensifying with‌ the increasing complexity of‌ methods, particularly with the‌‌ emergence of deep neural networks, generative AI, and‌ large language models, which‌ are difficult to explain‌‌ and interpret. Empirical evaluation is essential, in particular‌ because of the complexity‌ of the algorithms and‌‌ the unpredictable nature of the data.

The approach‌ taken in ths pillar‌ is that of organizing‌‌ scientific competitions (also called “challenges”). Scientific competitions systematize‌ large-scale experiments and show‌ the effectiveness of participants‌‌ in solving complex problems. Annual competitions, organized on‌ the Codalab competition platform,‌ address various scientific or‌‌ industrial questions, evaluating the automatic algorithms submitted by‌ participants. The newer version‌ of Codalab, called Codabench‌‌, extends the capabilities of Codalab to benchmarks.‌

Both challenges and benchmarks‌ are crucial for comparing‌‌ models and understanding their behavior. Recent applications include:‌ improving decision-making, particularly useful‌ in fields like finance‌‌ and medicine; helping to combat climate change by‌ optimizing the use of‌ resources; personalizing the customer‌‌ experience in e-commerce, banking, and other industries; improving‌ security and preventing fraud;‌ and improving accessibility for‌‌ people with disabilities, for example, through voice recognition‌ systems, visual aids for‌ the visually impaired, and‌‌ other assistive technologies.

The importance of impartial evaluations‌ of algorithms is constantly‌ increasing with the acceleration‌‌ of progress in Artificial Intelligence. According to David‌ Donoho: “The emergence of‌ Frictionless Reproducibility flows from‌‌ 3 data science principles that matured together after‌ decades of work by‌ many technologists and numerous‌‌ research communities. The mature principles involve data sharing,‌ code sharing, and competitive‌ challenges, however implemented in‌‌ the particularly strong form of frictionless open services.”‌ He cites the Codalab‌ project as being exemplary‌‌ in this area 92.

4 Application domains‌

4.1 Computational Social Sciences‌

Computational Social Sciences (CSS)‌‌ studies social and economic phenomena, ranging from technological‌ innovation to politics, from‌ media to social networks,‌‌ from human resources to education, from inequalities to‌ health. It combines perspectives‌ from different scientific disciplines,‌‌ building upon the tradition of computer simulation and‌ modeling of complex social‌ systems 96 on the‌‌ one hand, and data science on the other‌ hand, fueled by the‌ capacity to collect and‌‌ analyze massive amounts of digital data.

The emerging‌ field of CSS raises‌ formidable challenges along three‌‌ dimensions. Firstly, the definition of the research questions,‌ the formulation of hypotheses‌ and the validation of‌‌ the results require a tight pluridisciplinary interaction and‌ dialogue between researchers from‌ different backgrounds. Secondly, the‌‌ development of CSS is a touchstone for ethical‌ AI. On the one‌ hand, CSS gains ground‌‌ in major, data-rich private companies; on the other‌ hand, public researchers around‌ the world are engaging‌‌ in an effort to use it for the‌ benefit of society as‌ a whole 104.‌‌ The key technical difficulties relate to data and‌ model biases, and to‌ self-fulfilling prophecies. Thirdly, CSS‌‌ does not only regard‌ scientists: it is essential that the civil society‌ participate in the science of society 120.‌

Tao/TAU was involved in CSS for the‌ last five years, and its activities had been‌ strengthened thanks to P. Tubaro's and I. Guyon's‌ expertises respectively in sociology and economics, and in‌ causal modeling. Their departures has negatively impacted the‌ team activities in this domain, but many projects‌ are still on-going and CSS remains a domain‌ of choice (see Section 8.6).

4.2 Energy‌ Management

Energy Management has been an application domain‌ of choice for Tao since the mid 2000s,‌ with main partners SME Artelys (METIS Ilab Inria;‌ ADEME projects POST and NEXT), RTE (three CIFRE‌ PhDs), and IFPEN (bilateral contract, DATAIA project ML4CFD).‌ The goals concern i) optimal planning over several‌ spatio-temporal scales, from investments on continental Europe/North Africa‌ grid at the decade scale (POST), to daily‌ planning of local or regional power networks (NEXT);‌ ii) monitoring and control of the French grid‌ enforcing the prevention of power breaks (RTE); iii)‌ improvement of house-made numerical methods using data-intense learning‌ in all aspects of IFPEN activities (Section 8.4.2‌).

The daily maintainance of power grids requires‌ the building of approximate predictive models on the‌ top of any given network topology. Deep Networks‌ are natural candidates for such modelling, considering the‌ size of the French grid ( $\sim$ 10000‌ nodes), but the representation of the topology is‌ a challenge when, e.g. the RTE goal is‌ to quickly ensure the "n-1" security constraint (the‌ network should remain safe even if any of‌ the 10000 nodes fails). Existing simulators are too‌ slow to be used in real time, and‌ the size of actual grids makes it intractable‌ to train surrogate models for all possible (n-1)‌ topologies (see Section 8.5 for more details).

Another‌ aspect of Power Grid management regards the real-time‌ control of the grid topology, man-made at the‌ moment. Its automation is yet a difficult challenge,‌ but results on the L2RPN challenge have demonstrated‌ its feasibility with Reinforcement Learning, opening the way‌ to more ambitious goals (e.g., decentralized control via‌ multi-agent Reinforcement Learning, see Section 8.5).

4.3‌ Data-driven Numerical Modeling

In domains where both first‌ principle-based models and equations, and empirical or simulated‌ data are available, their combined usage can support‌ more accurate modelling and prediction, and when appropriate,‌ optimization, control and design, and help improving the‌ time-to-design chain through fast interactions between the simulation,‌ optimization, control and design stages. The expected advances‌ regard: i) the quality of the models or‌ simulators (through data assimilation, e.g. coupling first principles‌ and data, or repairing/extending closed-form models); ii) the‌ exploitation of data derived from different distributions and/or‌ related phenomenons; and, most interestingly, iii) the task‌ of optimal design and the assessment of the‌ resulting designs.

A first challenge regards the design‌ of the model space, and the architecture used‌ to enforce the known domain properties (symmetries, invariance operators, temporal structures). When‌ appropriate, data from different‌ distributions (e.g. simulated vs‌‌ real-world data) will be reconciled, for instance taking‌ inspiration from real-valued non-volume‌ preserving transformations 88 in‌‌ order to preserve the natural interpretation.

Another challenge‌ regards the validation of‌ the models and solutions‌‌ of the optimal design problems. The more flexible‌ the models, the more‌ intensive the validation must‌‌ be. Along this way, generative models will be‌ used to support the‌ design of "what if"‌‌ scenarios, to enhance anomaly detection and monitoring via‌ refined likelihood criteria.

In‌ the application domains described‌‌ by Partial Differential Equations (PDEs), the goal of‌ incorporating machine learning into‌ classical simulators is to‌‌ speed up the simulations while maintaining as much‌ as possible the accuracy‌ ad physical relevance of‌‌ the proposed solutions. Many possible tracks are possible‌ for this; one can‌ build surrogate models, either‌‌ of the whole system, or of its most‌ computationaly costly parts; one‌ can search to provide‌‌ better initialization heuristics to numerical solvers, which make‌ sure that physical constraints‌ are satisfied. Or one‌‌ can inject physical knowledge/constraints at different stages of‌ the numerical solver.

5‌ Social and environmental responsibility‌‌

5.1 Footprint of research activities

The Laboratory (LISN)‌ is currently actively re-thinking‌ its carbon footprint, being‌‌ part of the Labo1.5 initiative. We participate in‌ working groups about GreenAI‌ (being able to measure,‌‌ compare and mitigate the negative impact of training‌ and inference for large‌ models). To start changing‌‌ practices, the simple fact of reporting the cost‌ of training one's model‌ in publications has been‌‌ spotted as en efficient tool. Ideally, the development‌ cost (all the trainings‌ performed during the research,‌‌ not just the training of the model presented‌ in the paper) should‌ also be mentionned.

Another‌‌ axis studied by the lab is the limitation‌ of (aerial) transport, keeping‌ in mind that the‌‌ younger members should be allowed to build their‌ own research network and‌ foreign experiences.

5.2 Impact‌‌ of research results

All our work on Energy‌ (see Sections 4.2)‌ is ultimately targeted toward‌‌ optimizing the distribution of electricity, be it in‌ planning the investments in‌ the power network by‌‌ more accurate previsions of user consumption, or helping‌ the operators of RTE‌ to maintain the French‌‌ Grid in optimal conditions.

A collaboration with IDEEV‌ has just started, with‌ the idea of leveraging‌‌ Deep Learning as a tool to help unlock‌ agro-ecological research. In particular,‌ we aim to help‌‌ measure the yields in mixed cropping (requiring to‌ be able to classify‌ grains of a given‌‌ species but of different varieties – something impossible‌ to the naked eye)‌ and detect pollinators on‌‌ video footage taken outside (including wind, change in‌ light conditions, etc).

6‌ Highlights of the year‌‌

6.1 Awards

CNRS bronze medal for Flora Jay‌ (see link)‌
Best poster award for‌‌ Luca Teodorescu, Sorbonne Graduate conference on Machine learning‌ and Statistics 2025, November‌ 2025 (junior conference)

6.2‌‌ Prestigous Publications

In 2025,‌ the team has successfully submitted papers in the‌ most prestigious ML venues:

Neurips 2025: 42
ICML‌ 2025: 29
ICLR 2025 38, 28,‌ 27, 37
AAAI 2025 30

6.3 Visibility‌

Organization of Neurips EEG Foundation Challenge by Bruno‌ Aristimunha et Sylvain Chevallier; largest competition organized by‌ academic in NeurIPS and biggest competition to date‌ on Codabench plateform.
SIGEVO keynote for Marc Schoenauer‌ at ACM-GECCO 2025, Malaga, Spain.
Invited talk at‌ Cerisy for Michèle Sebag to the colloque on‌ "DE LA DÉ-COÏNCIDENCE À LA « VRAIE VIE‌ »

7 Latest software developments, platforms, open data‌

7.1 Latest software developments

7.1.1 Codalab

Keywords:
Benchmarking,‌ Competition
Functional Description:

Challenges in machine learning and‌ data science are competitions running over several weeks‌ or months to resolve problems using provided datasets‌ or simulated environments. Challenges can be thought of‌ as crowdsourcing, benchmarking, and communication tools. They have‌ been used for decades to test and compare‌ competing solutions in machine learning in a fair‌ and controlled way, to eliminate “inventor-evaluator" bias, and‌ to stimulate the scientific community while promoting reproducible‌ science. Current production infrastructure has been consolidated in‌ 2021 (sovereign distributed storage, 20 GPU workers) thanks‌ to the sponsorship of Région Ile-de-France, ANR, Université‌ Paris-Saclay, CNRS, INRIA, and ChaLearn, to support 20,000‌ new users (2024), organizing or participating each year‌ to hundreds of competitions. Some of the areas‌ in which Codalab is used include Computer vision‌ and medical image analysis, natural language processing, time‌ series prediction, causality, and automatic machine learning. Codalab‌ has been selected by the Région Ile de‌ France to organize industry-scale challenges. Codalab has been‌ ranked first on scientific criteria, in an independent‌ international study: https://mlcontests.com/state-of-competitive-machine-learning-2023/. TAU continues expanding Codalab to‌ accommodate new needs, including teaching. Link to the‌ historical server (read-only) https://competitions.codalab.org.

@article{codalab_competitions_JMLR, author = {Adrien‌ Pavao and Isabelle Guyon and Anne-Catherine Letournel and‌ Dinh-Tuan Tran and Xavier Baro and Hugo Jair‌ Escalante and Sergio Escalera and Tyler Thomas and‌ Zhen Xu}, title = {CodaLab Competitions: An Open‌ Source Platform to Organize Scientific Challenges}, journal =‌ {Journal of Machine Learning Research}, year = {2023},‌ volume = {24}, number = {198}, pages =‌ {1–6}, url = {http://jmlr.org/papers/v24/21-1436.html} }
URL:
http://competitions.codalab.org
Contact:‌
Isabelle Guyon

7.1.2 Cartolabe

Name:
Cartolabe
Keyword:
Information‌ visualization
Functional Description:

The goal of Cartolabe is‌ to build a visual map representing the scientific‌ activity of an institution/university/domain from published articles and‌ reports. Using the HAL Database, Cartolabe provides the‌ user with a map of the thematics, authors,‌ and articles. ML techniques are used for dimensionality‌ reduction, cluster, and topic identification, visualization techniques are‌ used for a scalable 2D representation of the‌ results.

Cartolabe has, in particular, been applied to‌ the Grand Debat dataset (3M individual propositions from‌ French Citizen, see https://cartolabe.fr/map/debat). The results were used‌ to test both the scaling capabilities of Cartolabe‌ and its flexibility to non-scientific and non-English corpora.‌ We also added sub-map capabilities to display the result of a year/lab/word‌ filtering as an online‌ generated heatmap with only‌‌ the filtered points to facilitate the exploration. Cartolabe‌ has also been applied‌ in 2020 to the‌‌ COVID-19 Kaggle publication dataset (Cartolabe-COVID project) to explore‌ these publications.
URL:
http://www.cartolabe.fr/‌
Publication:
hal-02499006
Contact:
Philippe‌‌ Caillou
Participant:
5 anonymous participants
Partners:
LRI -‌ Laboratoire de Recherche en‌ Informatique, CNRS

7.1.3 DeepHyper‌‌

Keywords:
Deep learning, Autotuning, HPC
Functional Description:
Machine‌ learning algorithms are continually‌ evolving to serve diverse‌‌ applications, yet their development often entails a significant‌ trial-and-error process to identify‌ optimal learning pipelines.This is‌‌ compounded by the multitude of data preprocessing techniques,‌ prediction (or generative) models,‌ and learning procedures available,‌‌ each offering a range of configurable parameters, also‌ referred to as hyperparameters.‌ DeepHyper addresses this challenge‌‌ by automating the selection and configuration of algorithms‌ and their corresponding hyperparameters,‌ facilitating a streamlined approach‌‌ for engineers and scientists to comprehend and optimize‌ the learning pipeline. At‌ its core, DeepHyper employs‌‌ parallel Bayesian optimization, validated through rigorous testing involving‌ up to 8,000 parallel‌ tasks. This methodology is‌‌ adaptable for both single and multi-objective tasks, enabling‌ efficient early discarding of‌ costly training steps. Furthermore,‌‌ DeepHyper seamlessly integrates with various parallel backends, including‌ multi-threading, multi-processing, Clouds (via‌ the Ray library), and‌‌ MPI-based schedulers on supercomputers, enhancing its scalability and‌ versatility across different computing‌ environments. The development of‌‌ DeepHyper is supported by the TAU-team through advances‌ in learning theory for‌ improving and explaining its‌‌ core algorithms.
URL:
https://github.com/deephyper/deephyper
Contact:
Romain Egele

7.1.4‌ OmniPrint

Keyword:
Open data‌
Functional Description:

Benchmarks and‌‌ shared datasets have been fostering progress in deep‌ learning. While there is‌ an increasing number of‌‌ available datasets, there is a need for larger‌ ones. However, collecting and‌ labeling data is time-consuming‌‌ and expensive, and systematically varying environmental conditions is‌ difficult and necessarily limited.‌ Therefore, resorting to artificially‌‌ generated data is helpful to drive fundamental research‌ in deep learning. OmniPrint‌ is geared to generating‌‌ an unlimited amount of printed characters.

Character images‌ provide excellent benchmarks for‌ deep learning problems because‌‌ of their relative simplicity and visual nature while‌ opening the door to‌ high-impact real-life applications. A‌‌ conjunction of technical features is required to meet‌ our specifications: pre-rasterization manipulation‌ of anchor points, post-rasterization‌‌ distortions, natural background and seamless blending, foreground filling,‌ anti-aliasing rendering, and importing‌ new fonts and styles.‌‌ Modern fonts such as TrueType or OpenType are‌ made of straight line‌ segments and quadratic Bezier‌‌ curves, connecting anchor points. Thus, it is easy‌ to modify characters by‌ moving anchor points. This‌‌ allows users to perform vectors-space pre-rasterization geometric transforms‌ (rotation, shear, etc.) as‌ well as distortions (e.g.,‌‌ modifying the length of ascenders of descenders) without‌ incurring aberrations due to‌ aliasing when transformations are‌‌ done in pixel space (post-rasterization).

The key technical‌ contributions include implementing transformations‌ and styles such as‌‌ elastic distortions, natural background, foreground filling, and so‌ on, selecting characters from‌ the Unicode standard to‌‌ form alphabets from more‌ than 20 languages around the world, further grouped‌ into partitions, to facilitate creating meta-learning tasks, identifying‌ fonts, implementing character rendering with a low-level FreeType‌ font rasterization engine, which enables direct manipulation of‌ anchor points, adding anti-aliasing rendering, implementing and optimizing‌ utility code to facilitate dataset formatting. To our‌ knowledge, OmniPrint is the pioneering text image synthesizer‌ geared toward ML research, supporting pre-rasterization transforms, which‌ allows Omniprint to imitate handwritten characters to some‌ degree. More details can be found in the‌ paper (https://openreview.net/forum?id=R07XwJPmgpl, https://arxiv.org/abs/2201.06648).
URL:
https://github.com/SunHaozhe/OmniPrint
Contact:
Haozhe Sun‌

7.1.5 codabench

Keywords:
Competition, Benchmarking
Functional Description:

Obtaining‌ standardized crowdsourced benchmark of computational methods is a‌ major issue in data science communities. Dedicated frameworks‌ enabling fair benchmarking in a unified environment are‌ yet to be developed. Here we introduce Codabench,‌ an open-source, community-driven platform for benchmarking algorithms or‌ software agents versus datasets or tasks. Codabench, released‌ in summer 2023, is the follower of Codalab,‌ enabling the same features and more: inverted data‌ challenges, better user experience, easier platform administration and‌ robustness. Competition design is backward compatible allowing an‌ easy migration from Codalab to Codabench. A public‌ instance of Codabench (https://codabench.org) is open to everyone,‌ free of charge, and allows benchmark organizers to‌ compare fairly submissions, under the same setting (software,‌ hardware, data, algorithms), with custom protocols and data‌ formats. Codabench has unique features facilitating the organization‌ of benchmarks flexibly, easily and reproducibly, such as‌ the possibility of re-using templates of benchmarks, and‌ supplying compute resources on-demand. In 2024, Codabench has‌ registered more than 10,000 new users and computed‌ near 80,000 participants submissions.

@article{codabench, title = {Codabench:‌ Flexible, easy-to-use, and reproducible meta-benchmark platform}, author =‌ {Zhen Xu and Sergio Escalera and Adrien Pavão‌ and Magali Richard and Wei-Wei Tu and Quanming‌ Yao and Huan Zhao and Isabelle Guyon}, journal‌ = {Patterns}, volume = {3}, number = {7},‌ pages = {100543}, year = {2022}, issn =‌ {2666-3899}, doi = {https://doi.org/10.1016/j.patter.2022.100543}, url = {https://www.sciencedirect.com/science/article/pii/S2666389922001465} }‌
URL:
https://www.codabench.org/
Contact:
Isabelle Guyon
Partner:
Région Île-de-France‌

7.1.6 pyriemann-qiskit

Keywords:
Quantum programming, Riemannian geometry, Symmetric‌ positive definite matrices
Functional Description:
Literature on quantum‌ computing suggests it may offer an advantage compared‌ with classical computing in terms of computational time‌ and outcomes, such as for pattern recognition or‌ when using limited training sets. Bulding on the‌ Qiskit library on quantum computing, pyriemann-qiskit implements a‌ wrapper around quantum-enhanced support vector classifiers (QSVCs) and‌ variational quantum classifiers (VQCs), to use quantum classification‌ with Riemannian geometry. It also introduces a quantum‌ version of the MDM algorithm, a classifier operating‌ on the manifold of symmetric positive definite matrices.‌
URL:
https://pyriemann-qiskit.readthedocs.io/en/latest/
Publication:
hal-04040814v1
Contact:
Sylvain Chevallier
Partner:‌
IBM

7.1.7 pyriemann

Keywords:
Riemannian geometry, Hermitian positive‌ definite matrices, Symmetric positive definite matrices
Functional Description:‌

Pyriemann is a Python machine learning package based‌ on scikit-learn API. It provides a high-level interface‌ for processing and classification of real (resp. complex)-valued‌ multivariate data through the Riemannian geometry of symmetric (resp. Hermitian) positive definite‌ (SPD) (resp. HPD) matrices.‌

pyRiemann aims at being‌‌ a generic package for multivariate data analysis but‌ has been designed around‌ biosignals (like EEG, MEG‌‌ or EMG) manipulation applied to brain-computer interface (BCI),‌ estimating covariance matrices from‌ multichannel time series, and‌‌ classifying them using the Riemannian geometry of SPD‌ matrices. It is widely‌ used in the scientific‌‌ community with more than one million download.
URL:‌
https://pyriemann.readthedocs.io
Contact:
Sylvain Chevallier‌

7.1.8 braindecode

Keywords:
Brain-Computer‌‌ Interface, Deep learning
Functional Description:
BrainDecode is an‌ open-source Python toolbox for‌ decoding raw electrophysiological brain‌‌ data with deep learning models. It includes dataset‌ fetchers, data preprocessing and‌ visualization tools, as well‌‌ as implementations of several deep learning architectures and‌ data augmentations for analysis‌ of EEG, ECoG and‌‌ MEG. It is design for neuroscientists who want‌ to work with deep‌ learning and deep learning‌‌ researchers who want to work with neurophysiological data.‌
URL:
https://braindecode.org/stable/index.html
Contact:
Sylvain‌ Chevallier
Partner:
Roche

7.1.9‌‌ MOABB

Name:
Mother of all BCI Benchmarks
Keywords:‌
Brain-Computer Interface, Open data,‌ Benchmarking
Functional Description:
Mother‌‌ of all BCI Benchmarks (MOABB) allows to build‌ a comprehensive benchmark of‌ popular brain-computer interface algorithms‌‌ applied on an extensive list of freely available‌ EEG datasets. This is‌ an open science initiative,‌‌ serving as a reference point for the future‌ algorithmic developments. Build on‌ reference libraries like scikit-learn‌‌ and MNE-python, machine learning pipelines can be ranked‌ and promoted on a‌ website, providing a clear‌‌ picture of the different solutions available in the‌ field. This software has‌ 80k downloads and an‌‌ active international development community.
URL:
https://neurotechx.github.io/moabb/
Contact:
Sylvain‌ Chevallier

7.1.10 dnadna

Name:‌
Deep Neural Architectures for‌‌ DNA
Keywords:
Deep learning, Population genetics
Functional Description:‌
DNADNA provides utility functions‌ to improve development of‌‌ neural networks for population genetics and is currently‌ based on PyTorch. In‌ particular, it already implements‌‌ several neural networks that allow inferring demographic and‌ adaptive history from genetic‌ data. Pre-trained networks can‌‌ be used directly on real/simulated genetic polymorphism data‌ for prediction. Implemented networks‌ can also be optimized‌‌ based on user-specified training sets and/or tasks. Finally,‌ any user can implement‌ new architectures and tasks,‌‌ while benefiting from DNADNA input/output, network optimization, and‌ test environment.
URL:
https://mlgenetics.gitlab.io/dnadna/‌
Contact:
Flora Jay

7.2‌‌ New platforms

Participants: Isabelle Guyon, Anne-Catherine Letournel‌, Sylvain Chevallier,‌ Adrien Pavao.

CODALAB:‌‌ The TAU group is community lead (under the‌ leadership of Isabelle Guyon)‌ of the open-source Codalab‌‌ project, hosted by Université Paris-Saclay, whose goal‌ is to host competitions‌ and benchmarks in machine‌‌ learning 111. We have replaced the historical‌ server by a dedicated‌ server hosted in our‌‌ lab. Since inception in December 2021, over 40000‌ participants entered 640 public‌ competitions (see statistics).‌‌ The engineering team, overseen by Anne-Catherine Letournel (CNRS‌ engineer) includes two engineers‌ dedicated full time to‌‌ administering the platform and developing challenges: Adrien Pavao,‌ financed by a project‌ started in 2020 with‌‌ the Région Ile-de-France, et‌ Dinh-Tuan Tran, financed by the ANR AI chaire‌ of Isabelle Guyon, Ihsan Ullah, financed by a‌ collaboration with LBNL/CERN and IJCLAB, and Benjamn Bearce‌ financed by the ANR AI chaire of Isabelle‌ Guyon. Several other engineers are engaged as contractors‌ on a needs-be basis. The rapid growth in‌ usage led us to put in place a‌ new infrastructure. We have migrated the storage over‌ a distributed Minio (4 physical servers, each with‌ 12 disks of 16 TB) spread over 2‌ buildings for robustness, and added 10 more GPUs‌ to the existing 10 previous ones in the‌ backend. A lot of horsepower to suport Industry-strength‌ challenges, thanks for the sponsorship of région Ile-de-France,‌ ANR, Université Paris-Saclay, CNRS, Inria, and ChaLearn.
CODABENCH:‌ Codabench 122 is a new version of Codalab‌ emphasizing the orgnization of benchmarks, which can be‌ thought of as ever-lasting challenges, de-emphasizing competiton, and‌ favoring the comparison between algorithms. Codabench has also‌ all the capabilities of Codalab and will progressively‌ replace it. When Codabench is fully stable, we‌ will retire Codalab.

The V1 of Codabench was‌ launched in August 2023. The user base is‌ rapidly growing, and it was ranked first by‌ MLContest.com for organizing academic challenges, in front of‌ Kaggle (funded by Google) and Tianchi (funded by‌ Alibaba). There is more than 500 papers published‌ each year mentionning Codabench.

Sylvain Chevallier and Bruno‌ Aristimunha organized the largest NeurIPS 2025 challenge lead‌ by academic with 1200 partitipants; close to the‌ first one, Code Golf, organized by Google. This‌ challenge relies on unreleased data from the Child‌ Mind Institute in New York and deep learning‌ models submitted by participants were evaluated on GPU‌ from San Diego and Paris-Saclay computing centers.

7.3‌ Open data

8 New results

8.1 Frugal Learning‌

Participants: Guillaume Charpiat, Marc Schoenauer, Michèle‌ Sebag, Sylvain Chevallier, Alice Lacan,‌ François Landes, Bruno Aristimunha Pinto.

8.1.1‌ Neural Architecture Growth for Frugal AI

In Manon‌ Verbockhaven's PhD thesis, we studied how to optimally‌ grow a neural network architecture, to increase the‌ performance (in terms of loss) while keeping the‌ network as small as possible (in particular, avoiding‌ redundancy). We showed 54 how to formulate the‌ notion of "expressivity bottleneck" in an easily computable‌ manner, and obtain optimal neuron weights as the‌ result of a small SVD. We showed that‌ the approach can scale up, with an experiment‌ using ResNet18 on CIFAR-100.

Thanks to the European‌ project MANOLO and to an ENS grant (CDSN),‌ this work has been continued by new PhD‌ students, Styliani Douka and Théo Rudkiewicz, and two‌ post-docs, Stéphane Rivaud and Alex Davey. In particular,‌ we now allow the architecture to grow as‌ an arbitrary DAG (Directed Acyclic Graph) 31 (paper‌ accepted at ESANN 2025).

In 2025 we welcomed‌ 2 master interns, Félix Houdouin and Léo Burgund,‌ who worked on the theoretical optimization side of‌ the project, leading to a 2-step iterative process to estimate the best‌ neurons to add to‌ a layer (taking into‌‌ account the non-linearity of the activation function, contrarily‌ to previously) that is‌ very effective in practice,‌‌ and to the extension of our approach to‌ the case of Transformers‌ (self-attention head). Meanwhile, we‌‌ explored alternative ways to deal with convolutions, we‌ extended our approach to‌ the addition of ResNet‌‌ blocks (i.e. increasing depth), and explored the use‌ of information theory concepts‌ to spot expressivity bottlenecks.‌‌ We also started to consider Reinforcement Learning approaches‌ to develop growth strategies.‌ We maintain a library,‌‌ GroMo, that we will advertise soon.

8.1.2‌ Model Frugality

Within the‌ context of Alice Lacan‌‌ 's PhD 49, we applied generative modelling‌ for data augmentation in‌ transcriptomics 101. The‌‌ high computational requirements of mainstream generative models (GAN,‌ WGAN, diffusion model) for‌ such high dimensional domains‌‌ led to the design of a new frugal‌ generative modelling approach, based‌ on density alignment 100‌‌.

Within the context of Nicolas Atienza 's‌ PhD 48 (Cifre Thales,‌ co-supervised with Johanne Cohen,‌‌ LISN), the latent space of a teacher classifier‌ is decomposed using Information‌ Bottleneck principles. Specifically, a‌‌ new loss is used to train a student,‌ exploiting a low-dimensional mapping‌ built on the top‌‌ of the teacher' latent representation, with no loss‌ of accuracy.

8.1.3 Data‌ frugality

Frugal approaches regarding‌‌ data availability are especially relevant in the context‌ of brain signal analysis,‌ where the acquisition cost‌‌ could be prohibitive. We investigated the different approaches‌ to leverage information from‌ small datasets, that resulted‌‌ in an international workshop in the flagship conference‌ for brain-computer interface later‌ published in reference journal‌‌ for the community 23.

With the PhD‌ work of Thibault de‌ Surrel, we investigated how‌‌ to build anisotropic models of the brain signal‌ based on covaraince representation.‌ This allows to define‌‌ wrapped distribution 40 that could improve machine learning‌ models with a data‌ frugal approach 41.‌‌ This covariance-based representation of the signal called for‌ new visualization techniques to‌ better understand algorithms relying‌‌ on information geometry, that is published in TMLR‌ 22.

8.2 Toward‌ Good AI

Participants: Philippe‌‌ Caillou, Alessandro Leite, Marc Schoenauer,‌ Michèle Sebag, Sylvain‌ Chevallier, Flora Jay‌‌, Cyril Furtlehner, Guillaume Charpiat, Cyriaque‌ Rousselot, Nicolas Atienza‌, Haozhe Sun,‌‌ Shuyu Dong, Antoine Szatkownik, Burak Yelmen‌, Thi Tuyet Trang‌ Chau.

8.2.1 Causal‌‌ Learning

Causal learning is commonly regarded as a‌ key research direction to‌ enforce the properties of‌‌ good AI models in terms of explainability, verifiability‌ and fairness. Its importance‌ is acknowledged through the‌‌ PEPR-IA-Causalit-AI, started in 2024, gathering four French laboratories/teams‌ (Loria at Nancy; TAU‌ and CELESTE at UPSaclay;‌‌ LIG at Grenoble).

A result from Armand Lacombe's‌ PhD 102 was obtained‌ regarding the estimation of‌‌ the conditional average treatment effect, based on asymmetrical‌ latent representations 61.‌ Interestingly, this result offers‌‌ guarantees and comes with‌ a procedure to adjust its hyper-parameter.

In the‌ last year of Shuyu Dong's postdoc (partnership Fujitsu),‌ building upon previous results 89, 90,‌ we have tackled the notorious lack of scalability‌ of causal graph learning from observational data, focusing‌ on the linear setting. The proposed approach, called‌ DCILP, is a Divide-and-Conquer approach. Formally (divide phase),‌ sub-problems involving the Markov blanket of each variable‌ are defined and solved in parallel. The reconciliation‌ of these partial solutions (conquer phase) is formulated‌ as an integer linear programming (ILP) problem, and‌ delegated to a state-of-the-art ILP solver. The parallelization‌ of the divide phase and the use of‌ ILP solvers for the conquer phase results in‌ a very good trade-off between time-complexity and accuracy‌ 30. Shuyu now holds an AI research‌ fellow at LAMSADE, at PSL; our continued collaboration‌ focuses on ensuring the acyclicity of the eventual‌ causal graph. We showed that the acyclicity (DAGness)‌ property can be enforced through adding a linear‌ number of ILP variables and quadratic number of‌ constraints w.r.t. the number of causal variables (results‌ submitted in 2026).

In Th. Weikert's PhD, starting‌ in 2025 and co-supervised with Johanne Cohen (LISN)‌ and Marianne Clausel (Loria), we aim to extend‌ the divide phase in DCILP to the non-linear‌ setting. Current results show that high-order statistical moments‌ (chiefly skewness and kurtosis) can be used to‌ identify the primary causes in causal graphs for‌ some non-linear causal mechanisms. The extension of this‌ first result to the iterative identification of rank-2,‌ rank-3 etc, variables is on-going.

Audrey Poinsot's PhD‌ (Cifre Ekimetrics) defended her PhD in December 2025‌ 51, centered on creating and analyzing rigourous‌ bencmarking in causal context. Because data in that‌ area does not pertain to Big Data, the‌ PhD first focused on Causal Data Augmentation 112‌, as known causality links can be leveraged‌ to ease learning. But the lack of recognized‌ benchmarks in that area led us to propose‌ a comprehensive survey of deep structural causal models,‌ focusing on their ability to answer counterfactual queries‌ using observational data within known causal structures 113‌. The lessons learned lead to a position‌ paper at ICML 2025 35, and to‌ develop CausalProfiler 63, a synthetic benchmark generator‌ for Causal ML methods. Based on a set‌ of explicit design choices about the class of‌ causal models, queries, and data considered, the CausalProfiler‌ randomly samples causal models, data, queries, and ground‌ truths constituting the synthetic causal benchmarks.

8.2.2 Explainable‌ and Robust Learning

Nicolas Atienza's PhD (Cifre Thales,‌ co-supervised with Johanne Cohen, LISN; 1 patent Thales,‌ another patent pending) 48 is to tackle the‌ main three goals of Trustable AI, i.e. explainability,‌ reliability and frugality. After a first result toward‌ explainability combining multi-modal embeddings and Multi-Criteria Decision Aid‌69, we obtained a new result based‌ on Extreme Value Theory, offering guarantees against adversarial‌ examples under mild assumptions 27. Most interestingly, this approach is also‌ effective against out-of-distribution samples,‌ as shown theoretically and‌‌ experimentally.

Robustness and explainability also are at the‌ core of TAU participation‌ in the Inria Challenge‌‌ OceanIA, that started in 2021 117.‌ The goal is to‌ provide scientists with an‌‌ assistant, exploiting the TARA images to identify the‌ ecosystems in the diverse‌ sites of the data‌‌ collection. A main challenge is related to the‌ identification of outliers and‌ samples from novel classes‌‌ (unseen in the training data). The proposed approach‌ (article in preparation) proceeds‌ by examining and refining‌‌ the latent representation of a classifier in the‌ spirit of the Information‌ Bottleneck, building upon N.‌‌ Atienza's PhD 27. Also linked to OCeanIA,‌ TAU collaborated to a‌ new study of the‌‌ global air-sea CO2 fluxes by mapping in situ‌ CO2 fugacity with re-analysis‌ data of environmental predictor‌‌ variables 14.

During the last year of‌ the Horizon Europe project‌ TRUST-AI, and even though‌‌ our work within this project was completed, we‌ extended the results obtained‌ on Memetic Semantic Genetic‌‌ Programming (MSGP) 106. MSGP is able to‌ generate short, and hence‌ hopefully easily explainable expressions‌‌ for Symbolic Regression problems. We implemented a boosting‌ procedure around MSGP, which‌ improved the performances without‌‌ degrading too much the interpretability 18.

Isabelle‌ Guyon co-organized the NeurIPS'24‌ experiments to assess the‌‌ usefulness of LLMs as an Author Checklist Assistant‌ for Scientific Papers based‌ on the authors' feedback‌‌ for 234 papers submitted to that conference 60‌. The answers demonstrated‌ both the potential usefulness‌‌ for paper improvements and the risks about the‌ vulnerabilities of automatic reviewing.‌

Also, note that the‌‌ work described in Section 8.4.1 about Interpretable Learning‌ Effective Dynamics (iLED) framework‌ 19 also contributes to‌‌ this line of research, adding interpretability to the‌ Deep Learning approach to‌ dynamical systems simulations.

Along‌‌ similar lines, in a joint work with the‌ DATAFLOT team at LISN‌ 15 describes the use‌‌ of conformal prediction (CP), a finite-sample and distribution-free‌ technique for estimating prediction‌ intervals with marginal coverage‌‌ guarantees, to improve the interpretation of confidence intervals‌ given by Gaussian Processes‌ surrogate models in expensive‌‌ numerical simulations.

8.2.3 Improved Learning

Mathurin Videau is‌ a CIFRE PhD student‌ with Meta (that started‌‌ long before Meta had to obey Musk's worst‌ practices), co-supervised by Olivier‌ Teytaud (member of the‌‌ team until 2016). Mathurin's PhD focuses on post-hoc‌ improvements of fully trained‌ models by using black-box‌‌ algorithms to optimize an impactful but small part‌ of model. Using BBO‌ allows us to optimize‌‌ non-differentiable loss functions that match the user's true‌ goals more efficiently than‌ the loss used for‌‌ the initial training of the model by standard‌ backpropagation, loss that needs‌ to be differentiable 25‌‌. This allows for instance to exactly optimize‌ the Word Error Rate‌ in translation tasks, the‌‌ number of deaths in Doom RL agents, or‌ to let the user‌ interactively guide generative processes‌‌ by directly acting in‌ the latent space.

Mathurin also worked on assessing‌ the Mixture-of-Expert approach for computer vision, with rather‌ negative results 24: the benefits of using‌ an MoE architecture decrease when the model gets‌ larger. In the end, a simple linear router‌ performs best, suggesting that additional routing complexity yields‌ no consistent benefit.

Last but not least, Mathurin‌ and co-authors addressed the fixed tokenization issue in‌ LLMs by introducing an autoregressive U-Net that learns‌ to embed its own tokens as it trains‌ 42. Because tokenization now lives inside the‌ model, the same system can handle character-level tasks‌ and transfer knowledge across low-resource languages.

8.2.4 Towards‌ high-quality and private genomes based on generative neural‌ networks

In collaboration with the Institute of Genomics‌ of Tartu, we have been leveraging various types‌ of generative neural networks (Generative Adversarial Networks, diffusion‌ models and Restricted Boltzmann Machines) in order to‌ learn and sample the high dimensional distributions of‌ real genomic datasets 123. These artificial genomes‌ retain important characteristics of the real genomes (genetic‌ allele frequencies and linkage, hidden population structure, ...)‌ without copying them and have the potential to‌ be valuable assets in future genetic studies by‌ providing anonymous substitutes for private databases (such as‌ the ones hold by companies or public institutes‌ like the Institute of Genomics of Tartu).

One‌ important problem in this context is to asses‌ whether the generative model preserves the privacy of‌ sensitive data used for training. In 66 proposed‌ a new method called PRIVET to assess the‌ privacy of generated samples. Using extreme value statistics‌ on nearest-neighbor distances PRIVET is able to assigns‌ individual privacy leak scores to synthetic samples, reliably‌ detecting memorization and leakage—even in high-dimensional, low-sample, or‌ underfitting settings (e.g., genetic data), while current methods‌ only provide global, non-interpretable risk estimates.

8.3 Machine‌ Learning with/for Statistical Physics

Participants: Cyril Furtlehner,‌ François Landes, Sergio Chibbaro, Alena Shilova‌, Anaclara Alvez-Canepa, Nicolas Béreux, Nilo‌ Schwencke, Lorenzo Mensi, Cyriaque Rousselot,‌ Aurélien Decelle (EPM), Catania Giovanni (UCM),‌ Beatriz Seoane (UCM).

8.3.1 Energy based models‌

Generative models constitute an important piece of unsupervised‌ ML techniques, which is under rapid development. In‌ this context, insights from statistical physics are important,‌ especially for energy-based models such as restricted Boltzmann‌ machines. Over the years we have contributed to‌ build a global picture of the Restricted Boltzmann‌ Machine (RBM) and its learning process 84,‌ 85, 87, 86. The subject‌ of Nicolas Béreux's PhD is to build on‌ these works to develop effecient learning algorithms. In‌ 28 an efficient approach has been proposed which‌ rely on: (i) designing a pre-training strategy allowing‌ us to bypass the most severe 2nd order‌ phase transitions, based on the mapping between the‌ RBM and the Coulomb machine proposed in 86‌; (ii) introducing a novel framework for estimating‌ log-likelihood (LL) by leveraging the learning trajectory's softness, rather than relying on‌ temperature integration; (iii) setting‌ a variation of the‌‌ standard parallel tempering algorithm in which exchanges occur‌ between the parameters of‌ models trained at different‌‌ stages, rather than across temperatures thereby avoiding to‌ cross the first order‌ transition line.

Still in‌‌ the context of energy based models we have‌ proposed in 29 a‌ theoretical framework to analyze‌‌ overfitting. Using a simplified setting at first, we‌ show that finite data‌ corrections can be accurately‌‌ modeled through asymptotic random matrix theory calculations and‌ provide the counterpart of‌ generalized cross-validation in the‌‌ energy based model context. These findings offer strategies‌ to control overfitting in‌ discrete-variable models through empirical‌‌ shrinkage corrections, improving the management of overfitting in‌ energy-based generative models. Our‌ framework is shown to‌‌ hold in principle for arbitrary energy-based models by‌ deriving the neural tangent‌ kernel dynamics of the‌‌ score function under the score-matching algorithm.

Physics informed‌ machine learning is also‌ an important axis of‌‌ research in the team with several avenues. One‌ is concerned with learning‌ critical phenomena, i.e. phenomena‌‌ displaying specific scaling properties, where typically all scales‌ contribute. This is the‌ subject of Anaclara Alvez‌‌ thesis, investigating the question of how to exploit‌ scale invariance for processes‌ dispaying statistical self-similarity, like‌‌ avalanche processes with the objective of being able‌ to extrapolate the predictions‌ from small to large‌‌ scales. In 55 this question is adressed in‌ the context of neural‌ operator by fully characterizing‌‌ relevent inductive biases based on statistical equivariance for‌ linear operators in terms‌ of a Fourier-Mellin representation.‌‌ This leads us to propose Fourier-Mellin neural operators‌ as a possible solution‌ for extrapolation at large‌‌ scales.

8.3.2 Physics Informed Machine Learning

The second‌ avenue deals with physics‌ informed neural networks (PINN's)‌‌ 114, which is an appealing way to‌ solve PDE by inserting‌ the physics into the‌‌ loss function and which in principle is mesh-free.‌ In the context of‌ Nilo Schwencke's thesis we‌‌ address some of the shortcomings of this methods,‌ plagued in particular with‌ optimization problems. For this‌‌ we provide a certain number of contributions in‌ 38 by: (i) introducing‌ the idea of an‌‌ empirical tangent space—a fresh take on the learning‌ dynamics described by the‌ neural tangent kernel (NTK)‌‌ 98. This perspective gives rise to a‌ family of algorithms (ANaGRAM),‌ based on the various‌‌ possible projections of the functional gradient onto this‌ empirical tangent space; (ii)‌ identifying a key relation‌‌ between our formulation of Natural Gradient for PINNs‌ with the operators Green‌ function; (iii) providing an‌‌ efficient implementation of the simplest instantiation of Anagram,‌ with favorable scaling properties,‌ showing robustness and superior‌‌ empirical results to existing baseline; (iv) defining a‌ new, simple and principled‌ optimization criteria for the‌‌ collocation point problem, which is a direct byproduct‌ of our theoretical framework.‌ In an extension of‌‌ ANaGRAM 65, 45, an adaptive strategy‌ is designed to control‌ a crucial cutoff parameter,‌‌ helping to maintain a‌ feature learning regime until trigerring a kernel regime‌ in order to complete the training at a‌ target precision. This allows us in practice to‌ gain further several orders of magnitude in precision‌ and to optimize the problem to machine precision‌ for some linear PDE.

Another new line of‌ research in connection with the causal learning pillar‌ of the team has emerged, dealing with the‌ generalized response theory. In 58 we show how‌ this can be leveraged in practice, with state‌ of the art ML methods, to virtually perform‌ interventional causality experiments on stochastic linear and nonlinear‌ systems, possibly chaotic, using only observational data. Additionally‌ we provide some theoretical insights on this approach,‌ by computing the asymptotic efficiency of the linear‌ response based causal predictor in a case of‌ large scale Markov process network of linear interactions,‌ thanks to random matrix theory.

A place where‌ ML can help address fundamental physics questions is‌ the domain of glasses (how the structure of‌ glasses is related to their dynamics), which is‌ one of the major problems in modern theoretical‌ physics 68 and played a key role in‌ Giorgio Parisi's career (2021 Nobel prize laureate). In‌ 2024 we proved that rotation-equivariant GNNs outperform other‌ approaches in terms of generalization power, displaying especially‌ good generalization to unseen temperatures. Our approach was‌ benchmarked against other recent works in the roadmap‌ we participated in 16, confirming that it‌ is extremely promising; we currently are actively exploring‌ this avenue of research. The main PhD student‌ carrying out this research defended in 2024, and‌ a new PhD student, Luca Teodorescu, is following‌ up on these works.

A parallel line of‌ research consists in using replica computation (a tool‌ from statistical physics) to compute the whole statistics‌ of possibly learned models, in simple settings. This‌ has been used to show that the optimal‌ training imbalance is different from 0.5, in a‌ simplified setup of Anomaly Detection 34 .

8.4‌ Machine Learning for Numerical Simulations

8.4.1 Representations and‌ Reduced Order Models for Dynamical Systems

Participants: Marc‌ Schoenauer, Matthieu Nastorg, Emmanuel Menier,‌ Thibault Monsel, Lionel Mathelin (DATAFLOT team, LISN)‌, Onofrio Semeraro (DATAFLOT team, LISN), Theofanis‌ Ifaistos.

During his PhD thesis 110,‌ Emmanuel Menier also spent 3 months in Spring‌ 2023 in Prof. Petros Koumoutsakos' group at SEAS‌ - Harvard, John A. Paulson School of Engineering‌ and Applied Sciences. This fruitful collaboration ended up‌ proposing Interpretable Learning Effective Dynamics (iLED) framework, a‌ novel framework based on nonlinear dimension reduction thanks‌ to deep neural networks, that adds the benefit‌ of interpretability to previous Deep approaches 19.‌

In a same spirit but with different tools,‌ we investigated the use of the signature transform‌ as an encoder for learning non-Markovian dynamics in‌ a continuous-time setting 36. The signature transform‌ offers a continuous-time alternative with strong theoretical foundations‌ and proven efficiency in summarizing multidimensional information in time. We integrated a‌ signature-based encoding scheme into‌ encoder-decoder dynamics models and‌‌ demonstrated that it outperforms RNN-based alternatives in test‌ performance on synthetic benchmarks.‌ The code is available‌‌ here.

8.4.2 Foundational models for numerical simuations‌

Participants: Theofanis Ifaistos,‌ Guillaume Charpiat, Michele‌‌ Alessandro Bucci (Safran Tech), Marc Schoenauer,‌ Matthieu Nastorg, Emmanuel‌ Menier, Lionel Mathelin‌‌ (DATAFLOT team, LISN).

We investigate Large Foundational‌ Models for Numerical Simulations,‌ notably thanks to an‌‌ Action Exploratoire (PI Guillaume Charpiat) Large Physics Models‌. This AeX is‌ tightly linked with the‌‌ startup company AUGUR, founded by Emmanuel Menier,‌ Matthieu Nastorg and Alice‌ Lacan, all former PhD‌‌ students in the team. Matthieu Nastorg was first‌ hired on this AeX‌ for one year as‌‌ a post-doc and then moved to the start-up‌ company. A new PhD‌ student co-supervised by Augur,‌‌ namely Theofanis Ifaistos, was recruted on the AeX.‌ We are designing neural‌ architectures to handle simulation‌‌ meshes of varied geometries, resolutions and quality, and‌ hopefully varied physics as‌ well (e.g., such as‌‌ different Reynolds in the case of fluid dynamics).‌

8.4.3 Advances in sparse‌ recovery for inverse problem‌‌ and application in M/EEG

Participants: Matthieu Kowalski,‌ Dylan Sechet, Jean-Baptiste‌ Malagnoux, Pierre Barbault‌‌ (CentraleSupelec), Charles Soussen (CentraleSupelec), Christian-George Bénar‌ (Inserm), Bruno Torrésani‌ (Aix-Marseille Université).

Inverse‌‌ problems involve reconstructing underlying signals or images from‌ indirect or incomplete measurements‌ and often require additional‌‌ constraints or regularization to ensure unique and stable‌ solutions. Sparse coding addresses‌ these challenges by representing‌‌ signals with a small number of nonzero coefficients‌ in a suitable basis‌ or dictionary. Methods like‌‌ Convolutional Dictionary Learning build on this principle and‌ have been successfully applied‌ in areas such as‌‌ neuroimaging and audio signal analysis. All the subsequent‌ contributions were carried out‌ within the ANR BMWs‌‌ (Bayesian Meets Wavelets) project.

13 introduces LEMUR, an‌ Expectation-Maximization (EM) framework for‌ estimating Bernoulli-Gaussian model parameters‌‌ in inverse problems. By jointly estimating both the‌ signal and hyperparameters, LEMUR‌ reduces the need for‌‌ manual tuning and proves effective for structured inverse‌ problems, particularly with correlated‌ measurement operators.

39 revisits‌‌ the CHAMPAGNE algorithm from the Sparse Bayesian Learning‌ (SBL) literature and establishes‌ an explicit connection with‌‌ iteratively reweighted sparse coding. Beyond a reinterpretation, the‌ paper provides an algorithmic‌ view of SBL updates‌‌ as a reweighted ${ℓ}_{1}$ -type procedure, clarifying‌ why CHAMPAGNE promotes sparsity‌ through variance/scale parameters and‌‌ how this mechanism can be leveraged for practical‌ optimization.

32 introduces an‌ approach to solve inverse‌‌ problems by performing convolutive sparse decomposition (CDL) directly‌ in the sensor domain‌ before tackling the spatial‌‌ inverse step. The core message is that a‌ low-rank / sparse decomposition‌ carried out on the‌‌ measurements can yield results comparable to decompositions performed‌ in the source domain,‌ while potentially simplifying the‌‌ pipeline and improving practicality when the source-space decomposition‌ is costly or unstable.‌ This sensor-space perspective opens‌‌ the door to modular‌ inverse-problem solvers where structured time–frequency or convolutional representations‌ are extracted first, and the spatial mapping is‌ addressed second.

As an application-oriented contribution in the‌ M/EEG ecosystem of BMWs, 21 proposes to combine‌ Independent Component Analysis (ICA) with machine learning to‌ facilitate and standardize the interpretation of MEG recordings‌ in drug-resistant focal epilepsy. Using recordings from 41‌ patients, the authors train models that classify ICA‌ components from a set of 61 predefined features:‌ a first classifier separates artifacts from non-artifacts, and‌ a second stage distinguishes multiple physiological/clinical categories (including‌ epileptic components). This work targets a critical bottleneck‌ in MEG workflow and provides a data-driven route‌ to more reproducible pre-surgical assessment pipelines.

8.4.4 Machine‌ learning for evolutionary genomics

Participants: Guillaume Charpiat,‌ Flora Jay, Léo Planche, Arnaud Quelin‌.

Collaboration: Bioinfo Team (LISN), MNHN (Paris),‌ UNAM (Mexico), U Brown (USA), METU (Turkey).

These‌ past years, we have pursued our effort in‌ generative AI for genomics. One goal is to‌ help access precious genomic information from large-scale databases,‌ such as biobanks, that are often kept private‌ or hardly accessible. Generative models offer the opportunity‌ of creating realistic proxy genomes that can enrich‌ researchers' downstream analyses and help maintaining high population‌ diversity in scientific studies. Recently we focused on‌ (i) designing more frugal generative models that still‌ scale to long genomic sequences and (ii) developed‌ a privacy test to detect any privacy leakage‌ of a trained generative model 66. This‌ test is flexible and data-agnostic and we have‌ demonstrated its use for image and genomic data.‌

We have also worked on tools for demographic‌ inference. Our problematic is to solve the inverse‌ problem of reconstructing past evolutionary processes from the‌ genomic signal of contemporary or ancient individuals. First,‌ we evaluated tree inference methods that reconstruct genealogical‌ trees along the genome, which can be further‌ used for extracting information and later feed demographic/cultural‌ inference tools. We have shown that tree reconstruction‌ tools are biaised under certain condition 64.‌ Second, we implemented classical supervised ML tools to‌ infer migration and population size parameters from extracted‌ genomic information in a simulation-based likelihood-free setting 20‌.

Together with U Brown, we developed and‌ applied HMM and frequency-based approaches for detecting archaic‌ segments in modern and ancient human genomes and‌ discovered such genomic regions under selection 26.‌ Along this line, we keep analyzing paleogenetic data‌ in collaboration with ancient DNA wet labs (UNAM,‌ Mexico ; METU, Turkey) to shed light on‌ evolutionary history of phathogens 57 and humans 17‌.

We have more recently started studying antibiotic‌ resistance in bacteria with a first prepint on‌ imputing missing data with denoising entoencoders 59 and‌ ongoing work based on (i) RF, XGBoost and‌ MLP for nucleotide/amino acid motifs and (ii) genomic‌ LLMs on the original DNA sequence, for predicting‌ sensitivity and resistance to antibiotics.

8.5 Energy Management‌

Participants: Alessandro Leite (INSA Rouen), Marc Schoenauer, Eva Boguslawski,‌ Benjamin Donnot (RTE),‌ Matthieu Dussartre (RTE).‌‌

Our collaboration with RTE has a long history,‌ starting with Benjamin Donnot's‌ (2016-2019) 91 and Balthazar‌‌ Donon's 93 CIFRE PhDs, and is centered on‌ the maintainance of the‌ national French Power Grid.‌‌ Eva Boguslawski's CIFRE PhD with RTE, co-supervised by‌ Alessandro Leite and Marc‌ Schoenauer, started in Sept.‌‌ 2022, and will be defended in June 2026.‌ It addresses the control‌ of the grid through‌‌ decentralized decision process using multi-agent Reinforcement Learning, in‌ the line of the‌ LR2PN challenge that Eva‌‌ contributed to organize during her Master internship 118‌. During the second‌ year of her PhD,‌‌ she focused on the emulation of Zonal Controllers‌ for the Power System‌ Transport Problem 76.‌‌ Finally, she developped MARL2GRID-TR, the first multi-agent RL‌ (MARL) benchmark for power‌ grid topology and redispatching.‌‌ Built on RTE France’s high-fidelity simulation platform, this‌ benchmark supports decentrazed control‌ across substations and generators,‌‌ with configurable agent scopes, observability settings, expert-informed heuristics,‌ and safety-critical constraints. The‌ benchmark includes a suite‌‌ of realistic scenarios that expose key challenges, such‌ as coordination under partial‌ information, long-horizon objectives, and‌‌ adherence to hard physical constraints

8.6 Computational Social‌ Sciences

Participants: Philippe Caillou‌, Michèle Sebag,‌‌ Cyriaque Rousselot, Guillaume Bied, Soal Nathan‌.

8.6.1 Labor Studies‌

Participants: Philippe Caillou,‌‌ Michèle Sebag, Guillaume Bied, Solal Nathan‌, Bruno Crépon (ENSAE)‌.

Job markets The‌‌ DataIA project Vadore 73 (partners ENSAE and Pôle‌ Emploi/France Travail) benefits from‌ the sustained cooperation and‌‌ from the wealth of data gathered by France‌ Travail. The data management‌ is regulated along a‌‌ 3-partite convention (GENES-ENSAE, Univ Paris-Saclay, Pôle Emploi). Extensive‌ efforts have been required‌ to achieve the data‌‌ pipelines required to enable learning recommendation models and‌ exploiting them in a‌ confidentiality preserving way (G.‌‌ Bied's PhD, 70). The acceptability and performance‌ of the algorithm for‌ the job seekers has‌‌ been investigated using large-scale field experiments (100,000 job‌ seekers) in Feb. 2023‌ and June 2024.

The‌‌ learned models 74 have been successfully assessed along‌ several criteria. A first‌ criterion regards the performance‌‌ vs state of the art algorithms and the‌ current Pole Emploi recommender‌ system. A second criterion‌‌ regards the time-to-recommendation; circa 0.02 seconds are currently‌ required to deliver a‌ bunch of recommended job‌‌ offers to each job seeker.

A limitation of‌ these first experiments was‌ that job seekers received‌‌ only one mail with recommendations, which led to‌ a very small impact‌ on effective applications. In‌‌ 2025, a new experiment targeted flow recommendation, sending‌ recommendations to job seekers‌ each week. The 6‌‌ month experiment led to a significant increase of‌ job seekers applications compared‌ to control groups receiving‌‌ recommendations from France Travail previous algorithm.

An important‌ criterion, besides the performance‌ in terms of recall‌‌ and the time-to-solution, regards the fairness of the‌ recommendation model 72,‌ 71. A comprehensive‌‌ study examining gender-related gap‌ in several utilities (wages, types of contract, distance-to-job)‌ has been conducted, comparing the gaps observed in‌ actual hirings, in applications, and in recommendations. Interestingly,‌ the gap in recommendations closely mimics that in‌ actual hirings and in applications (if any, the‌ recommendation algorithm tends to decrease the gap).

Algorithmic‌ fairness in domains as sensitive as employment is‌ under scrutiny of French and European regulations. The‌ difficulty is to decouple the biases observed in‌ applications (thus reflecting job seekers' preferences, that should‌ be respected) from those due to recruiters (that‌ should not be perpetuated in the learned models).‌

An important limitation of job recommendations concerns the‌ congestion of the job market (share of job‌ offers paid attention to by job seekers). Recommender‌ systems tend to increase the congestion due to‌ the so-called popularity bias. Early attempts to prevent‌ the congestion have been investigated in 75,‌ using optimal transport.

Both fairness and congestion issues‌ are at the core of S. Nathan's PhD‌ (coll. Univ. Ghent, Belgium). In 2025, we proposed‌ a new recommendation algorithm taking into account both‌ the quality of the recommendation and minimizing the‌ number of jobs with too few recommendations33‌.

A key difficulty for research on ML-based‌ job recommendation is the lack of open and‌ representative datasets, owing to the very sensitive nature‌ of the data and the protection of vulnerable‌ persons. We collaborate with U. Gent, Belgium, welcoming‌ Guillaume Bied for his post-doc and Solal Nathan‌ for a PhD visit, on this topic.

8.6.2‌ Health and practices

Participants: Philippe Caillou, Michèle‌ Sebag, Armand Lacombe, Cyriaque Rousselot,‌ Olivier Allais (INRA), Julia Mink (Univ. Bonn,‌ DE), Florian Yger (INSA Rouen).

Continuing‌ our former partnership with INRAE (in the context‌ of the Initiative de Recherche Stratégique Nutriperso; 95‌), we proposed the HorapestDataIA project to‌ uncover the potential causal relationships between pesticide dissemination‌ and children's health (Cyriaque Rousselot's PhD). A major‌ bottleneck in environmental epidemiology is the lack of‌ direct pesticide exposure data. Existing sensor framework is‌ sparse (50 sensors nationwide in the CNEP campaign),‌ with extreme class imbalance (90% of measurements are‌ below detection limits), and annual purchase statistics (BNVD-S‌ dataset) often fail to capture the fine-grained temporal‌ dynamics required to study in-utero effects.

To address‌ this challenge, we developed a framework to construct‌ nationwide “virtual sensors”. This involved harmonizing heterogeneous data‌ sources (including sensor measurements (CNEP), agricultural land use‌ (RPG), purchase statistics (BNVD-S), and meteorological fields (ERA5-Land),‌ into spatial meshes 116.

The core contribution‌ of 2025 is the development of a deep‌ spatial model capable of predicting weekly pesticide concentrations.‌ The model is trained using shared representation across‌ 75 monitored substances to tackle the small available‌ dataset. Contextual features (crops, purchases) are processed using‌ CNN to capture fine-grained spatial structures and a‌ wind-aware mask is used to integrate weather information.‌ Evaluations using a leave-one-out sensor protocol demonstrated good generalization performance for the‌ model even in unseen‌ geographic context. It provides‌‌ significantly better prediction than the widely used purchased-based‌ proxies. This model has‌ been applied to generate‌‌ weekly exposure maps for all communes in Metropolitan‌ France. These high-resolution predictions‌ are now being integrated‌‌ with health data to assess health outcomes using‌ the SNDS (National Health‌ Data System) database.

8.6.3‌‌ Scientific Information System and Visual Querying

Participants: Philippe‌ Caillou, Michèle Sebag‌, Anne-Catherine Letournel,‌‌ Hande Gozukan, Jean-Daniel Fekete (AVIZ, Inria Saclay)‌.

A third area‌ of activity concerns the‌‌ 2D visualisation and querying of a corpus of‌ documents. Its initial motivation‌ was related to scientific‌‌ organisms, institutes or Universities, using their scientific production‌ (set of articles, authors,‌ title, abstract) as corpus.‌‌ The Cartolabe project (see also Section 7)‌ started as an Inria‌ ADT (coll. Tao and‌‌ AVIZ, 2015-2017). It received a grant from CNRS‌ (coll. Tau, AVIZ‌ and HCC-LRI, 2018-2019).

The‌‌ originality of the approach is to rely on‌ the content of the‌ documents (as opposed to,‌‌ e.g. the graph of co-authoring and citations). This‌ specificity allowed to extend‌ Cartolabe to various corpora,‌‌ such as Wikipedia, Bibliotheque Nationale de France, or‌ the Software Heritage. Cartolabe‌ was also applied in‌‌ 2019 to the Grand Debat dataset: to support‌ the interactive exploration of‌ the 3 million propositions;‌‌ and to check the consistency of the official‌ results of the Grand‌ Debat with the data.‌‌ Cartolabe has also been applied in 2020 to‌ the COVID-19 kaggle publication‌ dataset (Cartolabe-COVID project) to‌‌ explore these publications.

Among its intended functionalities are:‌ the visual assessment of‌ a domain and its‌‌ structuration (who is expert in a scientific domain,‌ how related are the‌ domains); the coverage of‌‌ an institute expertise relatively to the general expertise;‌ the evolution of domains‌ along time (identification of‌‌ rising topics). A round of interviews with beta-user‌ scientists has been performed‌ in 2019-2020. Cartolabe usage‌‌ raises questions at the crossroad of human-centered computing,‌ data visualization and machine‌ learning: i) how to‌‌ deal with stressed items (the 2D projection of‌ the item similarities poorly‌ reflects their similarities in‌‌ the high dimensional document space; ii) how to‌ customize the similarity and‌ exploit the users' feedback‌‌ about relevant neighborhoods. A statement of the current‌ state of the project‌ was published in 2021‌‌ 80.

8.7 Organization of Challenges

Participants: Isabelle‌ Guyon, Marc Schoenauer‌ , Anne-Catherine Letournel,‌‌ Sylvain Chevallier, Audrey Poinsot, Alessandro Leite‌, Lisheng Sun,‌ Bruno Aristimunha.

8.7.1‌‌ Challenges

The Tau group uses challenges (scientific competitions)‌ as a means of‌ stimulating research in machine‌‌ learning and engage a diverse community of engineers,‌ researchers, and students to‌ learn and contribute advancing‌‌ the state-of-the-art. The Tau group is community lead‌ of the open-source Codalab‌ platform (see Section 7‌‌), hosted by Université Paris-Saclay. Codabench, the‌ new version of Codalab,‌ was financed in 2021‌‌ by a 500k€ project‌ with the Région Ile-de-France. This project also received‌ the support of the Chaire Nationale d'Intelligence Artificielle‌ of Isabelle Guyon (2020-2024), Fair Universe project), and‌ TAILOR ICT48 Network of Excellence (2020-2024). In 2025,‌ only remained the funding of Lawrence Berkeley Labs‌ (2022-2025). Anne-Catherine Letournel, engineer from LISN, is still‌ working on the project.

A particular highlight this‌ year has been the EEG Foundation Challenge: From‌ Cross-Task to Cross-Subject EEG Decoding ().‌ Bruno Aristimunha was one of its main organizers,‌ while Isabelle Guyon and Sylvain Chevallier were on‌ the Strategic Board. This challenge was the largest‌ competition ever organized by academic in NeurIPS and‌ biggest competition to date on Codabench plateform.

8.7.2‌ Datasets

TAU also contributed this year to designing‌ several datasets in order to improve comparative experiments‌ or challenges in several domains.

Stylized Meta-Album (SMA)‌ 62 is a new image classification meta-dataset comprising‌ 24 datasets (12 content datasets, and 12 stylized‌ datasets), designed to advance studies on out-of-distribution (OOD)‌ generalization and related topics, co-authored by Lisheng Sun‌ and Isabelle Guyon.

Causal Profiler63 is a‌ synthetic benchmark generator for Causal ML methods (see‌ Section 8.2.1) developped by Audrey Poinsot during‌ her PhD 51.

9 Bilateral contracts and‌ grants with industry

Participants: Whole team.

9.1‌ Bilateral contracts with industry

Tau continues its policy‌ about technology transfer, accepting any informal meeting following‌ industrial requests for discussion (and we are happy‌ to be often solicited), and deciding about the‌ follow-up based upon the originality, feasibility and possible‌ impacts of the foreseen research directions, provided they‌ fit our general canvas. This lead to the‌ following 4 on-going CIFRE PhDs, with the corresponding‌ side-contracts with the industrial supervisor, and the continuation‌ until September 2023 of the bilateral contract with‌ Fujitsu (within the national "accord-cadre" Inria/Fujitsu).

CIFRE RTE‌ 2021-2025 (72 kEuros), with RTE, related to Eva‌ Boguslawski's CIFRE PhD Decentralized Partially Observable Markov Decision‌ Process for Power Grid Management

Coordinator: Marc Schoenauer‌ and Matthieu Dussartre (RTE)

Participants: Eva Boguslawski, Alessandro‌ Leite
CIFRE MAIR 2025-2028 (75 kEuros), with Meta‌ (Facebook) AI Research, related to Ismail Labiad CIFRE‌ PhD New Training Approaches for Large Language Models‌

Coordinator: Matthieu Kowalski, Marc Schoenauer and Julia Kempe‌ (Meta)
CIFRE Ekimetrics 2022-2025 (45 kEuros), with Ekimetrics,‌ related to Audrey Poinsot's CIFRE PhD Causal incertainty‌ quantification under partial knowledge and low data regimes‌

Coordinator: Marc Schoenauer and Nicolas Chesneau (Ekimetrics)

Participants:‌ Guillaume Charpiat, Alessandro Leite, Audrey Poinsot and Michèle‌ Sebag
CIFRE MAIR 2022-2025 (75 kEuros), with Meta‌ (Facebook) AI Research, related to Mathurin Videau's CIFRE‌ PhD Reinforcement Learning: Sparse Noisy Reward

Coordinator: Marc‌ Schoenauer and David Lopez-Paz (Meta)

Participants: Alessandro Leite‌ and Mathurin Videau
CIFRE MAIR 2022-2025 (75 kEuros),‌ with Meta (Facebook) AI Research, related to Badr‌ Youbi's CIFRE PhD Learning invariant representations from temporal‌ data

Coordinator: Michèle Sebag and David Lopez-Paz (Meta)‌

Participants: Badr Youbi
AI Verse, related to‌ Abir Affane's post-doc

Coordinator: Pierer Alliez (INRIA Titane)

Participant: Guillaume Charpiat

10‌ Partnerships and cooperations

10.1‌ European initiatives

10.1.1 Horizon‌‌ Europe

Adra-e

Participants: Marc Schoenauer.

Adra-e project‌ on cordis.europa.eu

Title:
AI,‌ Data and Robotics ecosystem‌‌
Duration:
From July 1, 2022 to June 30,‌ 2025
Partners:
- INSTITUT NATIONAL‌ DE RECHERCHE EN INFORMATIQUE‌‌ ET AUTOMATIQUE (INRIA), France
- LINKOPINGS UNIVERSITET (LIU), Sweden‌
- UNIVERSITY OF GALWAY (OLLSCOIL‌ NA GAILLIMHE), Ireland
- DUBLIN‌‌ CITY UNIVERSITY (DCU), Ireland
- AI DATA AND ROBOTICS‌ ASSOCIATION (ADRA), Belgium
- TRUST-IT‌ SERVICES SRL (Trust-IT), Italy‌‌
- COMMISSARIAT A L ENERGIE ATOMIQUE ET AUX ENERGIES‌ ALTERNATIVES (CEA), France
- UNIVERSITEIT‌ TWENTE (UNIVERSITEIT TWENTE), Netherlands‌‌
- DEUTSCHES FORSCHUNGSZENTRUM FUR KUNSTLICHE INTELLIGENZ GMBH (DFKI), Germany‌
- ATOS SPAIN SA, Spain‌
- HRVATSKA UDRUGA ZA UMJETNU‌‌ INTELIGENCIJU (CROATIAN ARTIFICIAL INTELLIGENCE ASSOCIATION), Croatia
- COMMPLA SRL‌ (Commpla Srl), Italy
- ATOS‌ IT SOLUTIONS AND SERVICES‌‌ IBERIA SL (ATOS IT), Spain
- SIEMENS AKTIENGESELLSCHAFT, Germany‌
- UNIVERSITEIT VAN AMSTERDAM (UvA),‌ Netherlands
Inria contact:
Joost‌‌ Gueurst (DPE)
Coordinator:
Marc Schoenauer
Summary:
AI, Data‌ and Robotics (ADR) is‌ omnipresent in our daily‌‌ lives and key to addressing some of the‌ most pressing challenges facing‌ our society. Europe has‌‌ excellent research centres, innovative start-ups, a world-leading position‌ in robotics and competitive‌ manufacturing and services sectors,‌‌ from automotive to healthcare, energy, financial services and‌ agriculture. While the essentials‌ are present, European ADR‌‌ is waiting for exploitation to achieve its full‌ potential. The ADR ecosystem‌ is inherently complex because‌‌ many stakeholders at many different levels require a‌ holistic strategy towards collaboration‌ to be effective and‌‌ efficient. The Adra Association, representing the private side‌ of the ADR Partnership,‌ leverages this diversity through‌‌ its founding organisations (BDVA, euRobotics, CLAIRE, ELLIS, EurAI)‌ and channels it to‌ the benefit of the‌‌ European ecosystem. The Adra-e CSA proposal is set‌ up in close liaison‌ with Adra Association and‌‌ includes it as a partner, committed to sustain‌ its outcomes. Adra-e should‌ be seen as the‌‌ operational arm of the partnership to foster collaboration,‌ convergence and interoperability between‌ communities and disciplines to‌‌ advance European ADR while safeguarding the interest of‌ European citizens. This is‌ achieved by supporting the‌‌ ADR Partnership in the update and implementation of‌ the SRIDA, creating the‌ conditions for an inclusive,‌‌ sustainable, effective, multi-layered, and coherent European ADR ecosystem,‌ leading to increased trust‌ and adoption of ADR,‌‌ a more competitive supply and demand sides in‌ the EU and raising‌ private investments at the‌‌ same time.The consortium is composed of leading industry‌ and research organisations with‌ significant expertise in all‌‌ three disciplines. All are involved in Adra and‌ the associations and partnerships‌ shaping European research. Many‌‌ of them are supporting the Digitising European Industry‌ initiative from the EC‌ participating in the constitution‌‌ of Digital Innovation Hubs Network and Digital platforms.‌

MANOLO

MANOLO project on‌ cordis.europa.eu

Title:
Trustworthy Efficient‌‌ AI for Cloud-Edge Computing
Duration:
From January 1,‌ 2024 to December 31,‌ 2026
Partners:
- INSTITUT NATIONAL‌‌ DE RECHERCHE EN INFORMATIQUE ET AUTOMATIQUE (INRIA), France‌
- PAL ROBOTICS SLU (PAL‌ ROBOTICS), Spain
- LAUREA-AMMATTIKORKEAKOULU OY‌‌ (LAUREA UNIVERSITY OF APPLIED‌ SCIENCES), Finland
- UNIVERSITY COLLEGE DUBLIN, NATIONAL UNIVERSITY OF‌ IRELAND, DUBLIN (NUID UCD), Ireland
- Q-PLAN INTERNATIONAL ADVISORS‌ PC (Q-PLAN INTERNATIONAL), Greece
- FRAUNHOFER GESELLSCHAFT ZUR FORDERUNG‌ DER ANGEWANDTEN FORSCHUNG EV (Fraunhofer), Germany
- BIT &‌ BRAIN TECHNOLOGIES SL (BIT & BRAIN TECHNOLOGIES), Spain‌
- YRKESHOGSKOLAN ARCADA AB (ARCADA UNIVERSITY OF APPLIED SCIENCES‌ LTD), Finland
- EIT DIGITAL SPAIN, Spain
- TECHNISCHE UNIVERSITAET‌ BRAUNSCHWEIG, Germany
- EVIDEN TECHNOLOGIES SRL, Romania
- FOUR DOT‌ INFINITY INFORMATION AND TELECOMMUNICATIONS SOLUTIONS PRIVATE COMPANY (FOUR‌ DOT INFINITY LYSEIS PLIROFORIKIS KAI EPIKOINONION IDIOTIKI KEFALAIOUCHIKI‌ ETAIREIA), Greece
- UNIVERSITAT POLITECNICA DE CATALUNYA (UPC), Spain‌
- "NATIONAL CENTER FOR SCIENTIFIC RESEARCH ""DEMOKRITOS""" ("NCSR ""D"""),‌ Greece
- UNIVERSITE PARIS-SACLAY, France
- ATOS IT SOLUTIONS AND‌ SERVICES IBERIA SL (ATOS IT), Spain
- KATHOLIEKE UNIVERSITEIT‌ LEUVEN (KU Leuven), Belgium
- 28DIGITAL, Belgium
- ARX NET‌ AE YPIRESIES KAI EPICHIRISIS DIADIKTYOU ANONIMI ETAIRIA (ARX.NET‌ S.A.), Greece
Inria contact:
Guillaume Charpiat
Coordinator:
Summary:‌

MANOLO will deliver a complete stack of trustworthy‌ algorithms and tools to help AI systems reach‌ better efficiency and seamless optimization in their operations,‌ resources and data required to train, deploy and‌ run high-quality and lighter AI models in both‌ centralised and cloud-edge distributed environments. It will push‌ the state of the art in the development‌ of a collection of complementary algorithms for training,‌ understanding, compressing and optimising machine learning models by‌ advancing research in the areas of: model compression,‌ meta-learning (few-shot learning), domain adaptation, frugal neural network‌ search and growth and neuromorphic models. Novel dynamic‌ algorithms for data/energy efficient and policy-compliance allocation of‌ AI tasks to assets and resources in the‌ cloud-edge continuum will be designed, allowing for trustworthy‌ widespread deployment.

To support these activities a data‌ management framework for distributed tracking of assets and‌ their provenance (data, models, algorithms) and a benchmark‌ system to monitor, evaluate and compare new AI‌ algorithms and model deployments will be developed. Trustworthiness‌ evaluation mechanisms will be embedded at its core‌ for explainability, robustness and security of models while‌ using the Z-Inspection methodology for TrustworthyAI assesment, helping‌ AI systems conform to the new AI Act‌ regulation.

MANOLO will be deployed as a toolset‌ and tested in lab environments via Use Cases‌ with different distributed AI paradigms within cloud-edge continuum‌ settings; it will be validated in verticals such‌ as health, manufacturing, and telecommunications aligned with ADRA‌ identified market opportunities, and with a granular set‌ of embedded devices covering robotics, smartphones, IoT as‌ well as using Neuromorphic chips. MANOLO will integrate‌ with ongoing projects at EU level developing the‌ next operating system for cloud-edge continuum, while promoting‌ its sustainability via the AI-on-demand platform and EU‌ portals.

10.2 National initiatives

10.2.1 ANR

PEPR IA‌ SAIF (400k€) Safe AI through Formal methods

Coordinator:‌ Caterina Urban (INRIA Antique)

Participant: Guillaume Charpiat.
PEPR‌ IA CAUSALI-T-AI (400k€) CAUSALIty Teams up with Artificial‌ Intelligence

Coordinator: Marianne Clausel (Université de Loraine)

Participant:‌ Michèle Sebag, Alessandro Leite.
RoDAPoG 2021-2025 (302k€) Robust‌ Deep learning for Artificial genomics and Population Genetics‌

Coordinator: Flora Jay,

Participants: Cyril Furtlehner, Guillaume Charpiat.
ANR Scalp 2025-2029 SCALe‌ invariance and multifractal Processing‌ in neural networks

Coordinator:‌‌ Cyril Furtlehner

Participants: Sergio Chibbaro, François Landes, Lionel‌ Mathelin and Guillaume Charpiat‌

Partners: Ladhyx (IPP) and‌‌ IMT Atlantique

10.2.2 Others

Inria Challenge OceanAI 2021-2025,‌ AI, Data, Models for‌ a Blue Economy

Coordinator:‌‌ Nayat Sanchez Pi (Inria Chile)

Participants: Marc Schoenauer,‌ Michèle Sebag and Shiyang‌
DATAIA YARN 2022-2025, Automatic‌‌ Processing of Messy Brain Data with Robust Methods‌ and Transfer Learning

Coordinator:‌ Sylvain Chevallier

Participants: Florent‌‌ Bouchard (L2S), Fredéric Pascal (L2S), Alexandre Gramfort (Meta),‌ Sara Sedlar
Fair Universe‌ 2022-2025, We received with‌‌ Lawrence Berkeley Labs a grant of 6.4 million‌ USD to develop benchmarks‌ in High Energy Physics‌‌ and implement them on Codabench. Colaboration with David‌ Rousseau of IJCLAB.

Coordinator:‌ Isabelle Guyon

Participants: David‌‌ Rousseau, Ragansu Chakkappai, Ihsan Ullah
Action Exploratoire 2024-2026,‌ Large Physics Models

Coordinator:‌ Guillaume Charpiat

Participants: Marc‌‌ Schoenauer, Matthieu Nastorg (post-doc), Theofanis Ifaistos (PhD student)‌

10.3 Regional initiatives

Working‌ group on ML and‌‌ Statistical physics on the Saclay plateau ()‌

11 Dissemination

All Members‌

11.1 Promoting scientific activities‌‌

Michèle Sebag, IA, du passé au futur. Académie‌ Territoriale des Savoirs en‌ Construction, Figeac, juin 2025.‌‌
Michèle Sebag, L’invasion de ChatGPT dans l’éducation :‌ pour, contre, et comment.‌ Blog binaire, 25 avril‌‌ 2025.

11.1.1 Scientific events: organisation

Matthieu Kowalski: Unrolling‌ and un/self/*/supervised learning for‌ inverse problems, journée‌‌ du GDR IASIS, Paris, May 2025
Flora Jay:‌ LEGEND 2nd edition of‌ the international conference on‌‌ Machine Learning for Evolutionary Genomic Data. Aussois, Dec‌ 8th-12th
Flora Jay: LEGO‌ workshop at GDR BIMMM‌‌ day, Nantes, Nov 27-28th.
Flora Jay: Scientific committee‌ for the CNRS conference‌ Prospective Biodiversity & AI‌‌ 2025 - organization of the AI & Genomes‌ workshop, Paris Oct 13-14th‌

Chair of conference program‌‌ committees

Flora Jay: Area chair for ISMB/ECCB 2025‌ Proceedings. International conference with‌ Proceedings. Intelligent Systems for‌‌ Molecular Biology, European Conference on Computational Biology

Reviewer‌

All TAU members are‌ reviewers of the main‌‌ conferences in their respective fields of expertise.

11.1.2‌ Journal

Member of the‌ editorial boards

Marc Schoenauer‌‌ - Advisory Board, Evolutionary Computation Journal, MIT Press,‌ and Genetic Programming and‌ Evolutionary Machines, Springer Verlag.‌‌
Michèle Sebag - Editorial Board, ACM Transactions on‌ Evolutionary Learning and Optimization.‌

Reviewer

All members of‌‌ the team reviewed numerous articles for the most‌ prestigious journals in their‌ respective fields of expertise.‌‌

11.1.3 Invited talks

François Landes, Building Good Representations‌ of Local Environments with‌ Rotation-Equivariant Neural Networks,‌‌ "Navigating Rugged Landscapes", Bangalore, India, Nov. 2025 (invited‌ by CEFIPRA)
François Landes,‌ Learning representations of glassy‌‌ liquids with roto-translation equivariant Graph Neural Networks,‌ UMR MIA, AgroParisTech, lab‌ seminar, Oct 2025
François‌‌ Landes, Active Learning Strategy for pollinator classification in‌ low-resolution video capture,‌ IDEEV, lab seminar, July‌‌ 2025
Marc Schoenauer, Evolutionary Computation: Back to the‌ Future, SIGEVO keynote‌ at ACM-GECCO 2025, Malaga,‌‌ July 18. 2025.
Cyril Furtlehner, Online feature learning‌ in terms of spectral‌ flow processes, Un‌‌ quart de siècle pour‌ un quart de plan, Imera Marseille, April 15-17‌ 2025.
Cyril Furtlehner, Training Restricted Boltzmann Machines despite‌ phases transitions panelist at the JETC, 26–30 May,‌ 2025, Belgrade, Serbia
Sergio Chibbaro, invited to Simons‌ Foundation Collaboration on Wave Turbulence Annual Meeting taking‌ place in NYC. This collaboration, directed by Jalal‌ Shatah of New York University, is the first‌ attempt for a systematic coordinated study of wave‌ turbulence theory in a large-scale project, bringing together‌ state-of-the-art skills in the areas of mathematics and‌ physics, with theoretical, experimental and numerical expertise.
Guillaume‌ Charpiat, invitation to organize a workshop day (5‌ Feb. 2025) on our works on Frugal AI‌ at the Erwin Schrödinger Institute in Vienna, within‌ the workshop “Infinite-dimensional Geometry: Theory and Applications”.
Guillaume‌ Charpiat, Neural Network Growth for Frugal AI,‌ 24 January 2025 at the SHARP PEPR IA‌ in Lyon + 14 February 2025 at the‌ Optimization for AI day at ENS Lyon
Guillaume‌ Charpiat, Deep Learning for Numerical Simulations: 2 approaches‌ for ML4CFD at DataIA Pitching Day on “AI‌ & Physics”, 8 September 2025, CentraleSupélec
Guillaume Charpiat,‌ Croissance de neurones pour l’IA frugale, at‌ Ecolab, Ministère de la Transition Écologique, 20 November‌ 2025
Flora Jay Workshop on generative IA for‌ biomedical and evolutionary applications, Queen Mary University of‌ London, Feb 25th 2025
Flora Jay Pitch at‌ DATAIA cluster inauguration day, April 7th 2025
Michele‌ Sebag, Pitch at ISC-PIF celebration 25th anniversary
Michele‌ Sebag, IA et dé-coïncidence, Colloque Cerisy (28 juin‌ 4 juillet 2025)

11.1.4 Leadership within the scientific‌ community

Sylvain Chevallier: President of the academic society‌ CORTICO, promoting the research in brain-computer interface; Executive‌ Committee, Institut de Convergence DataIA; Research Committee,‌ IA Cluster Paris-Saclay; Head of MSCA-Horizon Europe‌ Cofund DeMythif.AI
Flora Jay: member of GDR BiM‌ science board (23-now)
Marc Schoenauer: Advisory Board, ACM-SIGEVO,‌ Special Interest Group on Evolutionary Computation; Chair‌ of Advisory Board (Founding President 2015-2022), SPECIES, Society‌ for the Promotion of Evolutionary Computation In Europe‌ and Surroundings, that organizes the yearly series of‌ conferences EvoStar. Senior Reviewer, IJCAI.
Michèle Sebag: Member‌ of scientific council of the AMIES Labex; Area‌ Chair NeurIPS, ICML, ECML-PKDD; Senior Meta-Reviewer ECAI, nommée‌ membre d'honneur de la SIF.

11.1.5 Scientific expertise‌

Guillaume Charpiat: Jean Zay (GENCI/IDRIS) committee member for‌ resource allocation (GPU) demand expertise
Guillaume Charpiat and‌ Marc Schoenauer, "Comité de pilotage sur les graph-NN",‌ RTE
Sylvain Chevallier, "Conseil Scientifique", Inclusive Brain
Marc‌ Schoenauer, Scientific Advisory Board, BCAM, Bilbao, Spain
Marc‌ Schoenauer, "Conseil Scientifique", IFPEN
Marc Schoenauer, "Conseil Scientifique",‌ Mines Paritech
Marc Schoenauer, "Conseil Scientifique", ADEME
Marc‌ Schoenauer, "Conseil scientifique du numérique" (CSN) of "Direction‌ Générale des Finances Publiques" (DGFiP).
Michèle Sebag, member‌ of Jury FNRS (scientific proposals; promotion for researchers),‌ Belgium;
Michèle Sebag, member of Jury SNSF (scientific‌ proposals), Switzerland.
Michèle Sebag, hiring committee, Ecole Nationale‌ des Mines de St Etienne;
Michèle Sebag, Promotion‌ Committee, University College Dublin
Flora Jay, hiring commmittee,‌ MCF Univ. Paris-Saclay

11.1.6 Research administration

Guillaume Charpiat: head of the Data‌ Science department at LISN,‌ Université Paris-Saclay.
Michèle Sebag,‌‌ Member of Council: Institut Pascal, IRSN, ISC-PIF, AMIES.‌
Sylvain Chevallier, co-chair of‌ the Scientific Council of‌‌ Computer Science dept. from Universite Paris-Saclay; elected member‌ of executive committee of‌ University Institute

11.2 Teaching‌‌ - Supervision - Juries - Educational and pedagogical‌ outreach

11.2.1 Teaching

Licence‌ : Philippe Caillou, Computer‌‌ Science for students in Accounting and Management, 192h,‌ L1, IUT Sceaux, Univ.‌ Paris Sud.
Licence :‌‌ François Landes, Introduction to Statistical Learning, 51h, L3,‌ Univ. Paris-Sud.
Master :‌ François Landes, Foundational Principles‌‌ of Machine Learning, 25h, M1 Recherche (AI track),‌ U. Paris-Sud.
Master :‌ François Landes, Machine Learning,‌‌ 42h, M2 Recherche, Univ. Paris-sud, physics department (PCS‌ international Master)
Licence :‌ Matthieu Kowalski, Signal Processing,‌‌ L3, 25h, Univ. Paris-Saclay
Master : Guillaume Charpiat,‌ Deep Learning in Practice,‌ 24h, M2 Recherche, MVA‌‌ / Centrale-Supelec / MscIA
Master : Guillaume Charpiat,‌ Information Theory, 14h, M1‌ IA Paris-Sud.
Master: Sylvain‌‌ Chevallier, Machine learning algorithms, 12h, M1, Univ. Paris-Saclay.‌
Master : Isabelle Guyon,‌ Project: Creation of mini-challenges,‌‌ M2, Univ. Paris-Sud.
Master : Flora Jay, Population‌ genetics inference, 4h, M2,‌ U PSaclay.
Master: Matthieu‌‌ Kowalski, Signal Processing, 25h, M2, Univ. Paris-Saclay
Master:‌ Matthieu Kowalski, Sparse Coding,‌ 36h, M2, Univ. Paris-Saclay‌‌
Master : Michèle Sebag, Deep Learning, 4h; Reinforcement‌ Learning, 12h; M2 Recherche,‌ U. Paris-Sud.
Master :‌‌ Beatriz Seoane, Applied Statistics, 25h, M1 Recherche (AI‌ track), U. Paris-Saclay.
Continuing‌ education (ie teaching in‌‌ companies): Guillaume Charpiat, Machine Learning, Deep Learning and‌ Reinforcement Learning, 11.5 days.‌

11.2.2 Supervision

PhD Alice‌‌ LACAN, Génération de données transcriptomiques à l'aide de‌ modèles génératifs profonds49‌, Blaise Hanczar and‌‌ Michèle Sebag, defended 4/2/2025.
PhD Manon VERBOCKHAVEN, Spotting‌ expressivity bottlenecks in neural‌ networks and fixing them‌‌ by optimal architecture growth54, Sylvain Chevallier‌ and Guillaume Charpiat, defended‌ 28/03/2025.
PhD Nicolas ATIENZA,‌‌ Towards Reliable ML: Leveraging Multi-Modal Representations, Information Bottleneck‌ and Extreme Value Theory‌48, Michèle Sebag‌‌ and Johanne Cohen, defended 4/4/2025.
PhD Nilo SCHWENKE‌ Gradients naturel et méthodes‌ à noyaux pour les‌‌ réseaux de neurones informés par la physique (PINNs)‌52, Cyril Furtlehner,‌ defended 8/12/2025.
PhD Jean-Baptiste‌‌ MALAGNOUX, Apprentissage de dictionnaires convolutifs et factorisation non‌ négative de matrices pour‌ la séparation de sources‌‌ et les problèmes inverses50, Matthieu Kowalski,‌ defended 15/12/2025.
PhD Antoine‌ SZATKOWNIK, Modélisation générative dans‌‌ un espace latent et évalutations de données synthétiques‌ en génétique des population‌53, Flora Jay,‌‌ Burak Yelmen, Cyril Furtlehner and Guillaume Charpiat, defended‌ 15/12/2025.
PhD Audrey POINSOT,‌ Causal Uncertainty Quantification under‌‌ Partial Knowledge and Low Data Regimes51,‌ Nicolas Chesneau (Ekimetrics), Guillaume‌ Charpiat, Alessandro Leite, and‌‌ Marc Schoenauer, defended 16/12/2025.
PhD in progress -‌ Anaclara ALVEZ, Scale-Equivariant Neural‌ Networks from 1/11/2023, François‌‌ Landes and Cyril Furtlehner.
PhD in progress -‌ Dylan SECHET, Séparation de‌ sources audio de musique‌‌ à l'aide de réseau de neurones interprétables from‌ 1/10/2025, Matthieu Kowalski and‌ Marc Evrard (LISN)
PhD‌‌ in progress - Jad‌ YEHYA, Détection d'événements rares avec modélisation de motifs‌ locaux from 1/04/2025, Matthieu Kowalski and Thomas Moreau‌ (Inria Mind)
PhD in progress - Ismail LABIAD,‌ Nouvelles stratégies d'apprentissage pour les grands modèles de‌ langage from 3/03/2025, Matthieu Kowalski and Marc Schoenauer‌ and Julia Kempe (Meta)
PhD in progress -‌ Bruno ARISTIMUNHA PINTO, Deep learning for decoding electroencephalography‌ from 01/06/2023, Raphael Y de Camargo (UFABC Brazil),‌ Marie-Constance Corsi (Inria Nerv), Sylvain Chevallier
PhD in‌ progress - Nicolas BÉREUX, Interpretability and pattern extraction‌ in Restricted Boltzmann Machines from 1/11/2023, Beatriz Seoane‌ Bartolome, Cyril Furtlehner.
PhD in progress - Eva‌ BOGUSLAWSKI, Congestion handling on Power Grid governed by‌ complex automata, from 1/05/22, Alessandro Leite, Mathieu‌ Dussartre (RTE) and Marc Schoenauer
PhD in progress‌ - Thibault DE SURREL, Learning context invariant representations‌ for EEG data, from 1/11/2023, Florian Yger‌ (ENSICaen), Fabien Lotte (Inria Potioc), Sylvain Chevallier
PhD‌ in progress - Styliani DOUKA Growth strategies for‌ neural architectures from 01/01/2024, Guillaume Charpiat and Sylvain‌ Chevallier
PhD in progress - Badr Youbi IDRISSI,‌ Learning an invariant representation through continuously evolving data‌, from 01/10/22, David Lopez-Paz (Meta) and Michèle‌ Sebag
PhD in progress - Theofanis IFAISTOS, Foundation‌ Models for Numerical Simulations, from 1/2/2025, Sergio‌ Chibbaro and Marc Schoenauer
PhD in progress -‌ Florent MICHEL, Deep Learning for Dictionary Learning,‌ from 1/10/2022, Matthieu Kowalski and Thomas Moreau (Inria‌ Mind)
PhD in progress - Solal NATHAN, Job‌ recommendation, AI Ethics and Optimal Transport., 1/1/2023,‌ Michèle Sebag and Philippe Caillou
PhD in progress‌ - Arnaud QUELIN, Infering Human population history with‌ approximated Bayesian computation and machine learning, from ancient‌ and recent genomes' polymorphism data, from 1/10/22,‌ Frédéric Austerlitz (MNHN), Flora Jay
PhD in progress‌ - Cyriaque ROUSSELOT, Spatio-temporal causal discovery – Application‌ to modeling pesticides impact, from 1/10/22, Philippe‌ Caillou
PhD in progress - Théo RUDKIEWICZ, Growing‌ neural networks for frugal learning from 01/10/2024, Guillaume‌ Charpiat and Sylvain Chevallier
PhD in progress -‌ Luca TEODORESCU, Rotation Equivariant Neural Networks for Glasses‌ from 1/10/2025, François Landes (with a dérogation to‌ be director without HDR).
PhD in progress -‌ Sebastien VELUT, Understanding and addressing within-user variability in‌ reactive and passive Brain-Computer Interfaces since 13/11/2023, Frédéric‌ Dehais (ISAE SupAero), Marie-Constance Corsi (Inria Nerv), Sylvain‌ Chevallier
PhD in progress, Mathurin VIDEAU, Reinforcement Learning‌ with sparse reward, from 01/10/2021, Alessandro Leite,‌ Marc Schoenauer and David Lopez-Paz (Meta).

11.2.3 Juries‌

François Landes: head of the M1 and M2‌ AI track selection comittee (M1 and M2 combined:‌ 1000+ applicants per year). Also head of the‌ scholarship short-listing comittee.

12 Scientific production

12.1 Major‌ publications

1 inproceedingsN.Nicolas Atienza, R.‌Roman Bresson, C.Cyriaque Rousselot, P.‌Philippe Caillou, J.Johanne Cohen, C.‌Christophe Labreuche and M.Michèle Sebag. Cutting‌ the Black Box: Conceptual Interpretation of a Deep‌ Neural Net with Multi-Modal Embeddings and Multi-Criteria Decision‌ Aid.Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence‌IJCAI-24, Thirty-Third International Joint‌ Conference on Artificial Intelligence‌‌Jeju, South KoreaInternational Joint Conferences on Artificial‌ Intelligence Organization2024,‌ 3669-3678HAL DOI
2‌‌ inproceedingsN.Nicolas Atienza, C.Christophe Labreuche‌, J.Johanne Cohen‌ and M.Michèle Sebag‌‌. Provably Safeguarding a Classifier from OOD and‌ Adversarial Samples: an Extreme‌ Value Theory Approach.‌‌Proc. ICLR'25ICLR 2025 - The Thirteenth International‌ Conference on Learning Representations‌Singapore (SG), SingaporeJanuary‌‌ 2025HAL
3 inproceedingsN.Nicolas Bereux,‌ A.Aurélien Decelle,‌ C.Cyril Furtlehner,‌‌ L.Lorenzo Rosset and B.Beatriz Seoane.‌ Fast training and sampling‌ of Restricted Boltzmann Machines‌‌.13th International Conference on Learning Representations -‌ ICLR 2025Singaour, Malaysia‌March 2025HAL
4‌‌ inproceedingsG.Guillaume Charpiat, N.Nicolas Girard‌, L.Loris Felardos‌ and Y.Yuliya Tarabalka‌‌. Input Similarity from the Neural Network Perspective‌.NeurIPS 2019 -‌ 33th Annual Conference on‌‌ Neural Information Processing SystemsVancouver, CanadaDecember 2019‌HAL
5 inproceedingsS.‌Shuyu Dong, M.‌‌Michèle Sebag, K.Kento Uemura, A.‌Akito Fujii, S.‌Shuang Chang, Y.‌‌Yusuke Koyanagi and K.Koji Maruhashi. DCDILP:‌ a distributed learning method‌ for large-scale causal structure‌‌ learning.Proc. AAAI 2025AAAI 25 -‌ The 39th Annual AAAI‌ Conference on Artificial Intelligence‌‌Philadelphia (PA), United StatesFebruary 2025HAL
6‌ articleS.Sylvain Gelly‌, M.Marc Schoenauer‌‌, M.Michèle Sebag, O.Olivier Teytaud‌, L.Levente Kocsis‌, D.David Silver‌‌ and C.Csaba Szepesvari. The Grand Challenge‌ of Computer Go: Monte‌ Carlo Tree Search and‌‌ Extensions.Communications- ACM5532012,‌ 106-113HAL
7 article‌D.Diviyan Kalainathan,‌‌ O.Olivier Goudet, I.Isabelle Guyon,‌ D.David Lopez-Paz and‌ M.Michèle Sebag.‌‌ Structural Agnostic Modeling: Adversarial Learning of Causal Graphs‌.Journal of Machine‌ Learning Research2022HAL‌‌
8 inproceedingsA.Apolline Mellot, A.Antoine‌ Collas, S.Sylvain‌ Chevallier, A.Alexandre‌‌ Gramfort and D.Denis Engemann. Geodesic optimization‌ for predictive shift adaptation‌ on EEG data.‌‌Proc. NeuIPS'24NeurIPS 2024 - 38th Conference on‌ Neural Information Processing Systems‌Vancouver, CanadaDecember 2024‌‌HAL
9 inproceedingsH.Herilalaina Rakotoarison, L.‌Louisot Milijaona, A.‌Andry Rasoanaivo, M.‌‌Michèle Sebag and M.Marc Schoenauer. Learning‌ Meta-features for AutoML.‌ICLR 2022 - International‌‌ Conference on Learning Representations (spotlight)Virtual, United States‌April 2022HAL
10‌ inproceedingsN.Nilo Schwencke‌‌ and C.Cyril Furtlehner. ANAGRAM: a natural‌ gradient relative to adapted‌ model for efficient PINNS‌‌ learning.In proceeding of ICLR 2025ICLR‌ 2025 - International Conference‌ on Learning Representations13th‌‌ International Conference on Learning Representations (ICLR 2025)Singapour,‌ MalaysiaApril 2025HAL‌
11 articleM.Manon‌‌ Verbockhaven, T.Théo Rudkiewicz, S.Sylvain‌ Chevallier and G.Guillaume‌ Charpiat. Growing Tiny‌‌ Networks: Spotting Expressivity Bottlenecks‌ and Fixing Them Optimally.Transactions on Machine‌ Learning Research JournalOctober 2024HAL
12 inproceedings‌M.Mathurin Videau, B.Badr Youbi Idrissi‌, A.Alessandro Leite, M.Marc Schoenauer‌, O.Olivier Teytaud and D.David Lopez-Paz‌. From Bytes to Ideas: Language Modeling with‌ Autoregressive U-Nets.NeurIPS 2025 - Advances in‌ Neural Information Processing SystemSan Diego (CA), United‌ StatesDecember 2025HAL

12.2 Publications of the‌ year

International journals

13 articleP.Pierre Barbault‌, M.Matthieu Kowalski and C.Charles Soussen‌. LEMUR: Latent EM Unsupervised Regression for Sparse‌ Inverse Problems.IEEE Transactions on Signal Processing‌732025, 2087-2098HAL DOI back to‌ text
14 articleT.-T.Thi-Tuyet-Trang Chau, A.‌ E.Anna E. Dudek, H.Hernando Ombao‌ and S.Sylvain Chevallier. Impact of Surface‌ Ocean CO2 Atlas (SOCAT) observing network extensions on‌ the quantification of global air-sea CO2 fluxes.‌Science of the Total Environment999October 2025‌, 180265In press. HAL DOI back to‌ text
15 articleE.Edgar Jaber, V.‌Vincent Blot, N.Nicolas Brunel, V.‌Vincent Chabridon, E.Emmauel Remy, B.‌Bertrand Iooss, D.Didier Lucor, M.‌Mathilde Mougeot and A.Alessandro Leite. Conformal‌ Approach to Gaussian Process Surrogate Evaluation with Coverage‌ Guarantees.Journal of Machine Learning for Modeling‌ and Computing632025, 37-68HAL‌DOI back to text
16 articleG.Gerhard‌ Jung, R. M.Rinske M Alkemade,‌ V.Victor Bapst, D.Daniele Coslovich,‌ L.Laura Filion, F. P.François P.‌ Landes, A.Andrea Liu, F. S.‌Francesco Saverio Pezzicoli, H.Hayato Shiba,‌ G.Giovanni Volpe, F.Francesco Zamponi,‌ L.Ludovic Berthier and G.Giulio Biroli.‌ Roadmap on machine learning glassy liquids.Nature‌ Reviews Physics722025, 91-104HAL‌DOI back to text
17 articleD.Dilek‌ Koptekin, A.Ayça Aydoğan, C.Cansu‌ Karamurat, N. E.N. Ezgi Altınışık,‌ K. B.Kıvılcım Başak Vural, D. D.‌D. Deniz Kazancı, A. K.Ayça Küçükakdağ‌ Doğu, D.Damla Kaptan, H. C.‌Hasan Can Gemici, E.Eren Yüncü,‌ H.Hannah Moots, G.Gülsün Umurtak,‌ R.Refik Duru, E.Erkan Fidan,‌ Ö.Özlem Çevik, B.Burçin Erdoğu,‌ T.Taner Korkut, C.Christopher Knüsel,‌ S.Scott Haddow, C. S.Clark Spencer‌ Larsen, R.Rana Özbal, F.Fokke‌ Gerritsen, E.Eylem Özdoğan, A.Ali‌ Akbaba, U. O.Uygar Ozan Usanmaz,‌ Y. C.Yasin Cemre Derici, M.Mine‌ Uçmazoğlu, F.Flora Jay, M.Mehmet‌ Özdoğan, A.Anders Götherström, Y. S.‌Yılmaz Selim Erdal, A.-S.Anna-Sapfo Malaspinas,‌ Ç.Çiğdem Atakuman, F.Füsun Özer and‌ M.Mehmet Somel. Out-of-Anatolia: cultural and genetic interactions during the Neolithic‌ expansion in the Aegean‌.Science3886754‌‌2025, adr3326HALDOI back to text‌
18 articleA.Alessandro‌ Leite and M.Marc‌‌ Schoenauer. Memetic Semantic Boosting for Symbolic Regression‌.Genetic Programming and‌ Evolvable Machines2611‌‌February 2025, 11HAL DOI back to‌ text
19 articleE.‌Emmanuel Menier, S.‌‌Sebastian Kaltenbach, M.Mouadh Yagoubi, M.‌Marc Schoenauer and P.‌Petros Koumoutsakos. Interpretable‌‌ learning of effective dynamics for multiscale systems.‌Proceedings of the Royal‌ Society A: Mathematical, Physical‌‌ and Engineering Sciences4812306January 2025,‌ 20240167HAL DOI back‌ to text back to‌‌ text
20 articleA.Arnaud Quelin, F.‌Frédéric Austerlitz and F.‌Flora Jay. Assessing‌‌ simulation-based supervised machine learning for demographic parameter inference‌ from genomic data.‌Heredity13472025‌‌, 417-426HAL DOIback to text
21‌ articleA.Aurore Semeux-Bernier‌, F.Francesca Bonini‌‌, S. M.Samuel Medina Villalon, M.‌Maria Fratello, M.‌Matthieu Kowalski, J.-M.‌‌Jean-Michel Badier, F.Frédéric Richard and C.-G.‌Christian-George Bénar. Classification‌ of magnetoencephalographic independent components‌‌ in epilepsy by machine learning.Clinical Neurophysiology‌1802025, 2111377‌HAL DOI back to‌‌ text
22 articleT.Thibault de Surrel,‌ S.Sylvain Chevallier,‌ F.Fabien Lotte and‌‌ F.Florian Yger. Geometry-Aware visualization of high‌ dimensional Symmetric Positive Definite‌ matrices.Transactions on‌‌ Machine Learning Research JournalFebruary 2025HAL back‌ to text
23 article‌M.Michael Tangermann,‌‌ S.Sylvain Chevallier, M.Matthias Dold,‌ P.Pierre Guetschel,‌ R.Reinmar Kobler,‌‌ T.Theodore Papadopoulo and J.Jordy Thielen.‌ Learning from small datasets‌ -- review of workshop‌‌ 6 of the 10th International BCI Meeting 2023‌.Journal of Neural‌ Engineering223June‌‌ 2025, 033001HALDOI back to text‌
24 article M.Mathurin‌ Videau, A.Alessandro‌‌ Leite, M.Marc Schoenauer and O.Olivier‌ Teytaud. Mixture of‌ Experts for Image Classification:‌‌ What's the Sweet Spot? Transactions on Machine Learning‌ Research Journal October 2025‌ HAL back to text‌‌
25 articleM.Mathurin Videau, M.Mariia‌ Zameshina, A.Alessandro‌ Leite, L.Laurent‌‌ Najman, M.Marc Schoenauer and O.Olivier‌ Teytaud. Evolutionary Retrofitting‌.ACM Transactions on‌‌ Evolutionary Learning and Optimization61February 2026‌, 1-45HAL DOI‌back to text
26‌‌ articleF. A.Fernando A Villanea, D.‌David Peede, E.‌ J.Eli J Kaufman‌‌, V.Valeria Añorve-Garibay, E. T.Elizabeth‌ T Chevy, V.‌Viridiana Villa-Islas, K.‌‌ E.Kelsey E Witt, R.Roberta Zeloni‌, D.Davide Marnetto‌, P.Priya Moorjani‌‌, F.Flora Jay, P. N.Paul‌ N Valdmanis, M.‌ C.María C Ávila-Arcos‌‌ and E.Emilia Huerta-Sánchez. The MUC19 gene:‌ An evolutionary history of‌ recurrent introgression and natural‌‌ selection.Science389‌6762August 2025, eadl0882HAL DOI back‌ to text

International peer-reviewed conferences

27 inproceedingsN.‌Nicolas Atienza, C.Christophe Labreuche, J.‌Johanne Cohen and M.Michèle Sebag. Provably‌ Safeguarding a Classifier from OOD and Adversarial Samples:‌ an Extreme Value Theory Approach.Proc. ICLR'25‌13th International Conference on Learning Representations - ICLR‌ 2025Singapore, SingaporeJanuary 2025HAL back to‌ text back to textback to text
28‌ inproceedingsN.Nicolas Béreux, A.Aurélien Decelle‌, C.Cyril Furtlehner, L.Lorenzo Rosset‌ and B.Beatriz Seoane. Fast training and‌ sampling of Restricted Boltzmann Machines.13th International‌ Conference on Learning Representations - ICLR 2025Singaour,‌ SingaporeMarch 2025HALback to text back‌ to text
29 inproceedingsG.Giovanni Catania,‌ A.Aurélien Decelle, C.Cyril Furtlehner and‌ B.Beatriz Seoane. A theoretical framework for‌ overfitting in energy-based modeling.ICML 2025 -‌ Forty-Second International Conference on Machine Learning267Vancouver,‌ CanadaPLMRJuly 2025, 6891-6919HAL back‌ to text back to text
30 inproceedingsS.‌Shuyu Dong, M.Michèle Sebag, K.‌Kento Uemura, A.Akito Fujii, S.‌Shuang Chang, Y.Yusuke Koyanagi and K.‌Koji Maruhashi. DCDILP: a distributed learning method‌ for large-scale causal structure learning.Proc. AAAI‌ 2025AAAI 25 - The 39th Annual AAAI‌ Conference on Artificial IntelligencePhiladelphia, United StatesFebruary‌ 2025HAL back to text back to text‌
31 inproceedingsS.Stella Douka, M.Manon‌ Verbockhaven, T.Théo Rudkiewicz, S.Stéphane‌ Rivaud, F. P.François P. Landes,‌ S.Sylvain Chevallier and G.Guillaume Charpiat.‌ Growth strategies for arbitrary DAG neural architectures.‌Proc. ESANN'25ESANN 2025 - 33th European Symposium‌ on Artificial Neural NetworksBruges, BelgiumApril 2025‌HAL DOI back to text
32 inproceedingsJ.-B.‌Jean-Baptiste Malagnoux and M.Matthieu Kowalski. Convolutive‌ Sparse Decomposition in Inverse Problems: A Sensor Space‌ Approach.SampTA 2025 - International Conference on‌ Sampling Theory and ApplicationsVienne, AustriaAugust 2025‌HAL DOI back to text
33 inproceedingsS.‌Solal Nathan, G.Guillaume Bied, E.‌Elia Pérennès, P.Philippe Caillou, B.‌Bruno Crépon, C.Christophe Gaillac and M.‌Michele Sebag. JoLA: Job Landscape Aware Job‌ Recommendation.RecSys in HR 2025 - The‌ 5th Workshop on Recommender Systems for Human Resources,‌ in conjunction with the 19th ACM Conference on‌ Recommender Systems4046Prague, Czech RepublicSeptember 2025‌HAL back to text
34 inproceedingsF.Francesco‌ Pezzicoli, V.Valentina Ros, F. P.‌François P. Landes and M.Marco Baity-Jesi.‌ Class Imbalance in Anomaly Detection: Learning from an‌ Exactly Solvable Model.AISTATS 2025 - 28th‌ International Conference on Artifi- cial Intelligence and Statistics‌Phuket, Thailand2025HALDOI back to text‌
35 inproceedingsA.Audrey Poinsot, P.Panayiotis‌ Panayiotou, A.Alessandro Leite, N.Nicolas Chesneau, Ö.Özgür‌ Şimşek and M.Marc‌ Schoenauer. Position: Causal‌‌ Machine Learning Requires Rigorous Synthetic Experiments for Broader‌ Adoption.Forty-second International‌ Conference on Machine Learning‌‌ Position Paper TrackVancouver, CanadaJuly 2025HAL‌back to text
36‌ inproceedingsE.Eliott Pradeleix‌‌, R.Rémy Hosseinkhan-Boucher, A.Alena Shilova‌, O.Onofrio Semeraro‌ and L.Lionel Mathelin‌‌. Learning Meets Differential Equations: From Theory to‌ Applications Learning non-Markovian Dynamical‌ Systems with Signature-based Encoders‌‌.https://proceedings.mlr.press/v277/pradeleix25a.htmlML-DE 2025 - 2nd Workshop on‌ "Machine Learning Meets Differential‌ Equations: From Theory to‌‌ Applications",277Bologna, ItalySeptember 2025, 1-25‌HAL back to text‌
37 inproceedingsS.Stéphane‌‌ Rivaud, L.Louis Fournier, T.Thomas‌ Pumir, E.Eugene‌ Belilovsky, M.Michael‌‌ Eickenberg and E.Edouard Oyallon. PETRA: Parallel‌ End-to-end Training with Reversible‌ Architectures.13th International‌‌ Conference on Learning Representations (ICLR 2025)The Thirteenth‌ International Conference on Learning‌ RepresentationsSingapour, Singapore2025‌‌HAL back to text
38 inproceedingsN.Nilo‌ Schwencke and C.Cyril‌ Furtlehner. ANAGRAM: a‌‌ natural gradient relative to adapted model for efficient‌ PINNS learning.In‌ proceeding of ICLR 2025‌‌13th International Conference on Learning Representations - ICLR‌ 202513th International Conference‌ on Learning Representations (ICLR‌‌ 2025)Singapour, SingaporeApril 2025HAL back to‌ text back to text‌
39 inproceedingsD.Dylan‌‌ Sechet, M.Matthieu Kowalski, S.Samy‌ Mokhtari and B.Bruno‌ Torrésani. Revisiting CHAMPAGNE:‌‌ Sparse Bayesian Learning as Reweighted Sparse Coding.‌SampTA 2025 - International‌ Conference on Sampling Theory‌‌ and ApplicationsVienna, AustriaJuly 2025HAL DOI‌back to text
40‌ inproceedingsT.Thibault de‌‌ Surrel, F.Fabien Lotte, S.Sylvain‌ Chevallier and F.Florian‌ Yger. Wrapped Gaussian‌‌ on the manifold of Symmetric Positive Definite Matrices‌.Proceedings of the‌ 42 nd International Conference‌‌ on Machine Learning, Vancouver, Canada. PMLR 267, 2025.‌ICML 2025 - 42nd‌ International Conference on Machine‌‌ LearningVancouver, CanadaJuly 2025HAL back to‌ text
41 inproceedingsT.‌Thibault de Surrel,‌‌ F.Florian Yger, F.Fabien Lotte and‌ S.Sylvain Chevallier.‌ A probabilistic view on‌‌ Riemannian machine learning models for SPD matrices.‌Lecture Notes in Computer‌ Science, vol 16035GSI‌‌ 2025 - 7th International Conference on Geometric Science‌ of InformationSaint Malo,‌ FranceOctober 2025,‌‌ 142–151HAL DOI back to text
42 inproceedings‌M.Mathurin Videau,‌ B.Badr Youbi Idrissi‌‌, A.Alessandro Leite, M.Marc Schoenauer‌, O.Olivier Teytaud‌ and D.David Lopez-Paz‌‌. From Bytes to Ideas: Language Modeling with‌ Autoregressive U-Nets.NeurIPS‌ 2025 - Advances in‌‌ Neural Information Processing SystemSan Diego (CA), United‌ StatesDecember 2025HAL‌back to text back‌‌ to text
43 inproceedingsB.Burak Yelmen,‌ M. N.Merve Nur‌ Güler, T.Tõnu‌‌ Kollo, M.Märt Möls, G.Guillaume‌ Charpiat and F.Flora‌ Jay. Bias in‌‌ genome-wide association test statistics‌ due to omitted interactions.RECOMB 2026 -‌ 30th Annual International Conference on Research in Computational‌ Molecular BiologyThessaloniki, GreeceNovember 2025HAL DOI‌

Conferences without proceedings

44 inproceedingsG.Guillaume Attuel‌ and M.Matthieu Kowalski. Disentangling Myth from‌ Reality: A Plea for exploring New Avenues.‌QIP25 - Quantum Information and Probability: from Foundations‌ to EngineeringVaxjö, SwedenJune 2025HAL
45‌ inproceedingsN.Nilo Schwencke, C.Cyriaque Rousselot‌, A.Alena Shilova and C.Cyril Furtlehner‌. AMSTRAMGRAM : Adaptive Multi-cutoff Strategy Modification for‌ ANaGRAM.ENUMATH 2025 - MS31 – Geometric‌ Optimization Methods for Scientific Machine LearningHeidelberg, Germany‌September 2025HAL back to text
46 inproceedings‌A.Alena Shilova, A.Alex Davey,‌ B.Brahim Driss and R.Riad Akrour.‌ StaQ it! Growing neural networks for Policy Mirror‌ Descent.EWRL — Eighteenth European Workshop on‌ Reinforcement LearningTubingen, Germany2025HAL DOI

Scientific‌ book chapters

47 inbookF.François Cabestaing and‌ S.Sylvain Chevallier. From signals to decisions‌ in noninvasive neural technologies.Neural InterfacesAcademic‌ Press; Elsevier2025, 77-90HAL DOI

Doctoral‌ dissertations and habilitation theses

48 thesisN.Nicolas‌ Atienza. Towards Reliable ML : Leveraging Multi-Modal‌ Representations, Information Bottleneck and Extreme Value Theory.‌Université Paris-SaclayApril 2025HAL back to text‌back to text back to text
49 thesis‌A.Alice Lacan. Transcriptomics data generation with‌ deep generative models.Université Paris-SaclayFebruary 2025‌HAL back to textback to text
50‌ thesisJ.-B.Jean-Baptiste Malagnoux. Convolutional dictionary learning‌ and non-negative matrix factorization for source separation and‌ inverse problems.Université Paris - SaclayDecember‌ 2025HAL back to text
51 thesisA.‌Audrey Poinsot. Structural Causal Models for Synthetic‌ Data Generation.Université Paris-SaclayDecember 2025HAL‌back to text back to text back to‌ text
52 thesisN.Nilo Schwencke. Natural‌ gradients and kernel methods for Physics Informed Neural‌ Networks (PINNs).Université paris saclayDecember 2025‌HAL back to text
53 thesisA.Antoine‌ Szatkownik. Latent generative modeling and synthetic data‌ evaluation in population genomics.Université Paris-SaclayDecember‌ 2025HAL back to text
54 thesisM.‌Manon Verbockhaven. Spotting expressivity bottlenecks in neural‌ networks and fixing them by optimal architecture growth‌.Université Paris-SaclayMarch 2025HAL back to‌ text back to text

Reports & preprints

55‌ miscA.Anaclara Alvez-Canepa, C.Cyril Furtlehner‌ and F.François Landes. Learning and extrapolating‌ scale-invariant processes.2026HAL DOI back to‌ text
56 miscG.Guillaume Attuel and M.‌Matthieu Kowalski. Disentangling myth from reality: Nonlocality‌ was not proven..February 2025HAL
57‌ miscM.Miriam Bravo-Lopez, E.Eduardo Arrieta-Donato‌, V.Viridiana Villa-Islas, Å. J.Åshild‌ Joanne Vågene, A.Ana Villaseñor-Altamirano, E.‌ G.Ernesto Garfias Morales, L.Laura Carrillo-Olivas‌, J.Jorge Gómez-Valdés, S. E.Sandra Elena Guevara Flores,‌ A. M.Alejandro Meraz‌ Moreno, M. M.‌‌Maria Moreno Cabrera, J.Jean Cury,‌ F.Flora Jay,‌ E.Emilia Huerta-Sánchez and‌‌ M.María Ávila-Arcos. Ancient genomic insights into‌ Salmonella enterica Paratyphi C‌ from Central Mexico.‌‌October 2025HAL DOIback to text
58‌ miscS.Sergio Chibbaro‌, C.Cyril Furtlehner‌‌, T.Théo Marchetta, A.-T.Andrei-Tiberiu Pantea‌ and D.Davide Rossetti‌. Building causation links‌‌ in stochastic nonlinear systems from data.September‌ 2025HAL DOI back‌ to text
59 misc‌‌E.Elsa Denakpo, N.Nicolas Dias,‌ J.Johann Pitout,‌ T.Thierry Naas,‌‌ D.Dylan Pillai, F.Flora Jay and‌ B. I.Bogdan I.‌ Iorga. Imputing missing‌‌ minimum inhibitory concentration (MIC) values for Pseudomonas aeruginosa‌ strains with a Denoising‌ AutoEncoder.March 2025‌‌HAL DOI back to text
60 miscA.‌Alexander Goldberg, I.‌Ihsan Ullah, T.‌‌ G.Thanh Gia Hieu Khuong, B. K.‌Benedictus Kent Rachmat,‌ Z.Zhen Xu,‌‌ I.Isabelle Guyon and N. B.Nihar B.‌ Shah. Usefulness of‌ LLMs as an Author‌‌ Checklist Assistant for Scientific Papers: NeurIPS’24 Experiment.‌September 2025HAL back‌ to text
61 misc‌‌A.Armand Lacombe and M.Michèle Sébag.‌ Asymmetrical Latent Representation for‌ Individual Treatment Effect Modeling‌‌.2025HAL DOIback to text
62‌ miscR.Romain Mussard‌, A.Aurélien Gauffre‌‌, I.Ihsan Ullah, T. G.Thanh‌ Gia Hieu Khuong,‌ M.-R.Massih-Reza Amini,‌‌ I.Isabelle Guyon and L.Lisheng Sun-Hosoya.‌ Stylized Meta-Album: Group-bias injection‌ with style transfer to‌‌ study robustness against distribution shifts.December 2025‌HAL back to text‌
63 miscP.Panayiotis‌‌ Panayiotou, A.Audrey Poinsot, A.Alessandro‌ Leite, N.Nicolas‌ Chesneau, M.Marc‌‌ Schoenauer and Ö.Özgür Şimşek. CausalProfiler: Generating‌ Synthetic Benchmarks for Rigorous‌ and Transparent Evaluation of‌‌ Causal Machine Learning.2025HAL DOI back‌ to text back to‌ text
64 miscF.‌‌Fanny Pouyet, F.Ferdinand Petit, J.‌Jérémy Guez, L.‌Léo Planche, E.‌‌Evelyne Heyer, B.Bruno Toupance, F.‌Flora Jay and F.‌Frédéric Austerlitz. Methods‌‌ for inferring coalescent tree topologies from genomic data:‌ a comparison based on‌ the transmission of reproductive‌‌ success.October 2025HAL back to text‌
65 miscN.Nilo‌ Schwencke, C.Cyriaque‌‌ Rousselot, A.Alena Shilova and C.Cyril‌ Furtlehner. AMStraMGRAM: Adaptive‌ Multi-cutoff Strategy Modification for‌‌ ANaGRAM.October 2025HAL back to text‌
66 miscA.Antoine‌ Szatkownik, A.Aurélien‌‌ Decelle, B.Beatriz Seoane, N.Nicolas‌ Béreux, L.Léo‌ Planche, G.Guillaume‌‌ Charpiat, B.Burak Yelmen, F.Flora‌ Jay and C.Cyril‌ Furtlehner. PRIVET: Privacy‌‌ Metric Based On Extreme Value Theory.October‌ 2025HAL back to‌ text back to text‌‌
67 miscJ.Jad‌ Yehya, M.Mansour Benbakoura, C.Cédric‌ Allain, B.Benoît Malezieux, M.Matthieu‌ Kowalski and T.Thomas Moreau. RoseCDL: Robust‌ and scalable convolutional dictionary learning for rare-event detection‌.2025HAL DOI

12.3 Cited publications

68‌ incollectionF.Francesco Arceri, F. P.François‌ P. Landes, L.Ludovic Berthier and G.‌Giulio Biroli. Glasses and aging: A Statistical‌ Mechanics Perspective.Encyclopedia of Complexity and Systems‌ Science (Living Reference)50 pages, 24 figs. This‌ is an updated version of a chapter initially‌ written in 2009 for the Encyclopedia of Complexity‌ and Systems Science (Springer)March 2022HAL DOI‌back to text
69 inproceedingsN.Nicolas Atienza‌, R.Roman Bresson, C.Cyriaque Rousselot‌, P.Philippe Caillou, J.Johanne Cohen‌, C.Christophe Labreuche and M.Michèle Sebag‌. Cutting the Black Box: Conceptual Interpretation of‌ a Deep Neural Net with Multi-Modal Embeddings and‌ Multi-Criteria Decision Aid.Proceedings of the Thirty-Third‌ International Joint Conference on Artificial IntelligenceJeju, South‌ KoreaInternational Joint Conferences on Artificial Intelligence Organization‌August 2024, 3669-3678HAL DOI back to‌ text
70 phdthesisG.Guillaume Bied. Designing‌ Recommender Systems for the Labor Market.Université‌ Paris-SaclayJuly 2024HALback to text
71‌ inproceedingsG.Guillaume Bied, C.Christophe Gaillac‌, M.Morgane Hoffmann, P.Philippe Caillou‌, B.Bruno Crépon, S.Solal Nathan‌ and M.Michele Sebag. Fairness in job‌ recommendations: estimating, explaining, and reducing gender gaps.‌Proceedings of the 1st Workshop on Fairness and‌ Bias in AI co-located with 26th European Conference‌ on Artificial Intelligence (ECAI 2023)3523Krakow, Poland‌CEUR-WS.orgOctober 2023HALback to text
72‌ inproceedingsG.Guillaume Bied, C.Christophe Gaillac‌, M.Morgane Hoffmann, S.Solal Nathan‌, P.Philippe Caillou, B.Bruno Crépon‌ and M.Michèle Sebag. Gender fairness in‌ job recommendation: a case study.AI for‌ HR and Public Employment ServicesGhent (BE), Belgium‌February 2023HAL back to text
73 inproceedings‌G.Guillaume Bied, S.Solal Nathan,‌ E.Elia Perennes, C.Christophe Gaillac,‌ P.Philippe Caillou, B.Bruno Crépon and‌ M.Michèle Sebag. Using Data from job‌ seekers, job offers and past hirings to learn‌ a Job Recommender System: the VADORE Project.‌AI for HR and Public Employment ServicesGhent‌ (BE), BelgiumFebruary 2023HAL back to text‌
74 inproceedingsG.Guillaume Bied, S.Solal‌ Nathan, E.Elia Perennes, M.Morgane‌ Hoffmann, P.Philippe Caillou, B.Bruno‌ Crépon, C.Christophe Gaillac and M.Michèle‌ Sebag. Toward Job Recommendation for All.‌IJCAI 2023 - The 32nd International Joint Conference‌ on Artificial IntelligenceMacau, ChinaInternational Joint Conferences‌ on Artificial Intelligence OrganizationAugust 2023, 5906-5914‌HAL DOI back to text
75 inproceedingsG.‌Guillaume Bied, E.Elia Perennes, S.Solal Nathan, V.‌Victor Naya, P.‌Philippe Caillou, B.‌‌Bruno Crépon, C.Christophe Gaillac and M.‌Michele Sebag. RECTO‌ : REcommandation diminuant la‌‌ Congestion par Transport Optimal.Proc. APIA 2023‌APIA2023AFIA and‌ ICubeStrasbourg, FranceAFIA‌‌July 2023, 89-98HAL back to text‌
76 inproceedingsE.Eva‌ Boguslawski, A.Alessandro‌‌ Leite, B.Benjamin Donnot, M.Matthieu‌ Dussartre and M.Marc‌ Schoenauer. Emulation of‌‌ Zonal Controllers for the Power System Transport Problem‌.ML4SPS 2024 -‌ Machine Learning for Sustainable‌‌ Power Systems ECML 2024 WorkshopVilnius, LithuaniaSeptember‌ 2024HAL back to‌ text
77 inproceedingsR.‌‌Roman Bresson, J.Johanne Cohen, E.‌Eyke Hüllermeier, C.‌Christophe Labreuche and M.‌‌Michele Sebag. Neural Representation and Learning of‌ Hierarchical 2-additive Choquet Integrals‌.IJCAI-PRICAI-20 - Twenty-Ninth‌‌ International Joint Conference on Artificial Intelligence and Seventeenth‌ Pacific Rim International Conference‌ on Artificial IntelligenceYokohama,‌‌ FranceJuly 2020, 1984-1991HAL DOI back‌ to text
78 inproceedings‌R.Roman Bresson,‌‌ J.Johanne Cohen, E.Eyke Hüllermeier,‌ C.Christophe Labreuche and‌ M.Michele Sebag.‌‌ On the Identifiability of Hierarchical Decision Models.‌18th International Conference on‌ Principles of Knowledge Representation‌‌ and Reasoning (KR-2021)Online, FranceInternational Joint Conferences‌ on Artificial Intelligence Organization‌November 2021, 151-162‌‌HAL DOI back to text
79 articleM.‌ M.Michael M Bronstein‌, J.Joan Bruna‌‌, T.Taco Cohen and P.Petar Veliċković‌. Geometric deep learning:‌ Grids, groups, graphs, geodesics,‌‌ and gauges.arXiv preprint arXiv:2104.134782021back‌ to text
80 article‌P.Philippe Caillou,‌‌ J.Jonas Renault, J.-D.Jean-Daniel Fekete,‌ A.-C.Anne-Catherine Letournel and‌ M.Michèle Sebag.‌‌ Cartolabe: A Web-Based Scalable Visualization of Large Document‌ Collections.IEEE Computer‌ Graphics and Applications41‌‌2April 2021, 76--88HAL DOI back‌ to text
81 inproceedings‌L.Lénaïc Chizat,‌‌ E.Edouard Oyallon and F.Francis Bach.‌ On Lazy Training in‌ Differentiable Programming.NeurIPS‌‌322019back to text
82 inproceedingsT.‌Taco Cohen and M.‌Max Welling. Group‌‌ Equivariant Convolutional Networks.Proc. ICML48PMLR‌2016, 2990--2999back‌ to text
83 misc‌‌G.Gwendoline De Bie, H.Herilalaina Rakotoarison‌, G.Gabriel Peyré‌ and M.Michèle Sebag‌‌. Distribution-Based Invariant Deep Networks for Learning Meta-Features‌.February 2021HAL‌back to text
84‌‌ articleA.A. Decelle, G.G. Fissore‌ and C.C. Furtlehner‌. Spectral dynamics of‌‌ learning in restricted Boltzmann machines.EPL (Europhysics‌ Letters)11962017‌, 60001back to‌‌ text
85 articleA.A. Decelle, G.‌G. Fissore and C.‌C. Furtlehner. Thermodynamics‌‌ of Restricted Boltzmann Machines and Related Learning Dynamics‌.J. Stat. Phys.‌1722018, 1576-1608‌‌back to text
86 articleA.A Decelle‌ and C.Cyril Furtlehner‌. Exact Training of‌‌ Restricted Boltzmann Machines on‌ Intrinsically Low Dimensional Data.Physical Review Letters‌September 2021HAL back to text back to‌ text
87 inproceedingsA.Aurélien Decelle, C.‌Cyril Furtlehner and B.Beatriz Seoane. Equilibrium‌ and non-Equilibrium regimes in the learning of Restricted‌ Boltzmann Machines.NeurIPS 2021Proceedings NeurIPS 2021‌Vancouver, United StatesDecember 2021HAL back to‌ text
88 inproceedingsL.Laurent Dinh, J.‌Jascha Sohl-Dickstein and S.Samy Bengio. Density‌ estimation using Real NVP.Int. Conf. on‌ Learning Representations (ICLR)2017back to text
89‌ inproceedingsS.Shuyu Dong and M.Michèle Sebag‌. From graphs to DAGs: a low-complexity model‌ and a scalable algorithm.ECML-PKDD 2022 -‌ European Conference on Machine Learning and Principles and‌ Practice of Knowledge Discovery in DatabasesGrenoble, France‌September 2022HAL back to text
90 unpublished‌S.Shuyu Dong, K.Kento Uemura,‌ A.Akito Fujii, S.Shuang Chang,‌ Y.Yusuke Koyanagi, K.Koji Maruhashi and‌ M.Michèle Sebag. Learning Large Causal Structures‌ from Inverse Covariance Matrix via Matrix Decomposition.‌October 2023, working paper or preprintHAL‌back to text
91 phdthesisB.Benjamin Donnot‌. Deep learning methods for predicting flows in‌ power grids : novel architectures and algorithms.‌Université Paris Saclay (COmUE)February 2019HAL back‌ to text
92 miscD.David Donoho.‌ Data Science at the Singularity.2023back‌ to text
93 phdthesisB.Balthazar Donon.‌ Deep statistical solvers & power systems applications.‌Université Paris-SaclayMarch 2022HAL back to text‌
94 articleJ.Jonathan Frankle and M.Michael‌ Carbin. The lottery ticket hypothesis: Finding sparse,‌ trainable neural networks.arXiv preprint arXiv:1803.036352018‌back to text
95 inproceedingsK.Ksenia Gasnikova‌, P.Philippe Caillou, O.Olivier Allais‌ and M.Michèle Sebag. Towards causal modeling‌ of nutritional outcomes.Causal Analysis Workshop Series‌ (CAWS) 2021519online, United States2021‌HAL back to text
96 articleJ.Jim‌ Giles. Computational social science: Making the links‌.Nature - News48874122012,‌ 448-450back to text
97 inproceedingsA.Aapo‌ Hyvarinen, H.Hiroaki Sasaki and R.Richard‌ Turner. Nonlinear ICA Using Auxiliary Variables and‌ Generalized Contrastive Learning.Proc. AISTAT89PMLR‌2019, 859--868URL: https://proceedings.mlr.press/v89/hyvarinen19a.htmlback to text‌
98 incollectionA.Arthur Jacot, F.Franck‌ Gabriel and C.Clement Hongler. Neural Tangent‌ Kernel: Convergence and Generalization in Neural Networks.‌Advances in Neural Information Processing Systems 312018‌, 8571--8580back to text
99 inproceedingsA.‌Arthur Jacot, F.Franck Gabriel and C.‌Clement Hongler. Neural Tangent Kernel: Convergence and‌ Generalization in Neural Networks.NeurIPS312018‌back to text
100 inproceedingsA.Alice Lacan‌, B.Blaise Hanczar and M.Michele Sebag‌. Frugal Generative Modeling for Tabular Data.‌Lecture Notes in Computer ScienceECML PKDD 2024 - European Conference on‌ Machine Learning and Principles‌ and Practice of Knowledge‌‌ Discovery in Databases14948Lecture Notes in Computer‌ ScienceVilnius, LithuaniaSpringer‌ Nature SwitzerlandAugust 2024‌‌, 55--72HAL DOIback to text
101‌ articleA.Alice Lacan‌, M.Michèle Sebag‌‌ and B.Blaise Hanczar. GAN-based data augmentation‌ for transcriptomics: survey and‌ comparative assessment.Bioinformatics‌‌39Supplement_1June 2023, i111-i120HAL DOI‌back to text
102‌ phdthesisA.Armand Lacombe‌‌. Changes of representation for counter-factual inference.‌Université Paris-SaclayMarch 2024‌HAL back to text‌‌
103 articleR.Remi Lam, A.Alvaro‌ Sanchez-Gonzalez, M.Matthew‌ Willson, P.Peter‌‌ Wirnsberger, M.Meire Fortunato, A.Alexander‌ Pritzel, S.Suman‌ Ravuri, T.Timo‌‌ Ewalds, F.Ferran Alet, Z.Zach‌ Eaton-Rosen, W.Weihua‌ Hu, A.Alexander‌‌ Merose, S.Stephan Hoyer, G.George‌ Holland, J.Jacklynn‌ Stott, O.Oriol‌‌ Vinyals, S.Shakir Mohamed and P.Peter‌ Battaglia. GraphCast: Learning‌ skillful medium-range global weather‌‌ forecasting.Science3826677cite arxiv:2212.127942022‌, 1416-142URL: http://arxiv.org/abs/2212.12794‌back to text
104‌‌ articleD.David Lazer, A.Alex Pentland‌, L.Lada Adamic‌, S.Sinan Aral‌‌, A. L.Albert L. Barabasi, D.‌Devon Brewer, N.‌Nicholas Christakis, N.‌‌Noshir Contractor, J.James Fowler, M.‌Myron Gutmann, T.‌Tony Jebara, G.‌‌Gary King, M.Michael Macy, D.‌Deb Roy and M.‌Marshall Van Alstyne.‌‌ Life in the network: the coming age of‌ computational social science.‌Science32359152009‌‌, 721–723back to text
105 inbookY.‌Yann Lecun, S.‌Sumit Chopra, R.‌‌Raia Hadsell, M. A.Marc Aurelio Ranzato‌ and F. J.Fu‌ Jie Huang. A‌‌ tutorial on energy-based learning.Predicting structured data‌G.G. Bakir,‌ T.T. Hofman,‌‌ B.B. Scholkopt, A.A. Smola and‌ B.B. Taskar,‌ eds. MIT Press2006‌‌back to text
106 inproceedingsA.Alessandro Leite‌ and M.Marc Schoenauer‌. Memetic Semantic Genetic‌‌ Programming for~Symbolic Regression.Lecture Notes in Computer‌ ScienceLNCS-13986Genetic Programming‌Species SocietyBrno, Czech‌‌ RepublicSpringer Nature SwitzerlandApril 2023, 198-212‌HAL DOI back to‌ text
107 inproceedingsM.‌‌Milad Leyli-Abadi, D.David Danan, M.‌Mouadh Yagoubi, S.-E.‌Seif-Eddine Attoui, A.‌‌Antoine Marot, J.J Picault, B.‌B Donnot, P.‌P Dimitrov, C.‌‌C Etienam and A.A Farjallah. LIPS‌ - Learning Industrial Physical‌ Simulation benchmark suite.‌‌NeurIPS - Data & Benchmark TrackNew Orleans,‌ United StatesNovember 2022‌HAL back to text‌‌
108 articleZ.Ziming Liu and M.Max‌ Tegmark. Machine Learning‌ Hidden Symmetries.Phys.‌‌ Rev. Lett.128182022, 180201back‌ to text
109 article‌C.Christos Louizos,‌‌ K.Karen Ullrich and‌ M.Max Welling. Bayesian compression for deep‌ learning.Advances in neural information processing systems‌302017back to text
110 thesisE.‌Emmanuel Menier. Deep Learning for Reduced Order‌ Modeling.Université Paris-SaclayJanuary 2024HAL back‌ to text
111 techreportA.Adrien Pavao,‌ I.Isabelle Guyon, A.-C.Anne-Catherine Letournel,‌ X.Xavier Baró, H.Hugo Escalante,‌ S.Sergio Escalera, T.Tyler Thomas and‌ Z.Zhen Xu. CodaLab Competitions: An open‌ source platform to organize scientific challenges.Université‌ Paris-Saclay, FRA.April 2022HAL back to text‌
112 unpublishedA.Audrey Poinsot and A.Alessandro‌ Leite. A Guide for Practical Use of‌ ADMG Causal Data Augmentation.March 2023,‌ Workshop on the pitfalls of limited data and‌ computation for Trustworthy ML, ICLR 2023, Kigali, Rwanda‌HAL back to text
113 inproceedingsA.Audrey‌ Poinsot, A.Alessandro Leite, N.Nicolas‌ Chesneau, M.Michele Sebag and M.Marc‌ Schoenauer. Learning Structural Causal Models through Deep‌ Generative Models: Methods, Guarantees, and Challenges.Proc.‌ Thirty-Third International Joint Conference on Artificial IntelligenceIJCAI‌ 2024 - Survey TrackProceedings of the Thirty-Third‌ International Joint Conference on Artificial IntelligenceJeju, South‌ KoreaAugust 2024, 8207-8215HAL DOI back‌ to text
114 articleM.Maziar Raissi,‌ P.Paris Perdikaris and G. E.George E‌ Karniadakis. Physics-informed neural networks: A deep learning‌ framework for solving forward and inverse problems involving‌ nonlinear partial differential equations.Journal of Computational‌ physics3782019, 686--707back to text‌
115 inproceedingsH.Herilalaina Rakotoarison, L.Louisot‌ Milijaona, A.Andry Rasoanaivo, M.Michèle‌ Sebag and M.Marc Schoenauer. Learning Meta-features‌ for AutoML.ICLR 2022 - International Conference‌ on Learning Representations (spotlight)Virtual, United StatesApril‌ 2022HAL back to text
116 inproceedingsC.‌Cyriaque Rousselot, O.Olivier Allais, P.‌Philippe Caillou, J.Julia Mink and F.‌Florian Yger. Assessing Impact of Pesticide Exposure‌ on Child Health using Large-Scale Data Integration and‌ Modelling.Workshop 2024 - Qualité de l'Air,‌ Agriculture et Santé HumaineSaint-Rémy-lès-Chevreuses, FranceNovember 2024‌HAL back to text
117 bookN.Nayat‌ Sanchez-Pi, L.Luis Marti, J.Julien‌ Salomon, J.Jacques Sainte-Marie, O.Olivier‌ Bernard, M.Michele Sebag, M.Marc‌ Schoenauer, A.Alejandro Maass, D.Damien‌ Eveillard, A.André Abreu, C.Colomban‌ de Vargas and P. A.Pablo A. Marquet‌. OcéanIA: AI, Data, and Models for Understanding‌ the Ocean and Climate Change.July 2021‌, 1-64HAL back to text
118 inproceedings‌G.Gaëtan Serré, E.Eva Boguslawski,‌ B.Benjamin Donnot, A.Adrien Pavão,‌ I.Isabelle Guyon and A.Antoine Marot.‌ Reinforcement learning for Energies of the future and‌ carbon neutrality: a Challenge Design.SSCI 2022‌ - IEEE Symposium Series on Computational IntelligenceIEEESingapour, SingaporeDecember 2022‌HAL back to text‌
119 inproceedingsC.Claudia‌‌ Shi, D. M.David M. Blei and‌ V.Victor Veitch.‌ Adapting Neural Networks for‌‌ the Estimation of Treatment Effects.NeurIPS 2019‌2019, 2503--2513URL:‌ https://proceedings.neurips.cc/paper/2019/hash/8fb5f8be2aa9d6c64a04e3ab9f63feee-Abstract.htmlback to text‌‌
120 bookI.I. Stengers. Another Science‌ Is Possible.Open‌ Humanities Press2013back‌‌ to text
121 inproceedingsY.Yixin Wang and‌ D.David Blei.‌ A Proxy Variable View‌‌ of Shared Confounding.Proc. 38th ICML139‌Proceedings of Machine Learning‌ ResearchPMLR2021,‌‌ 10697--10707URL: https://proceedings.mlr.press/v139/wang21c.htmlback to text
122 article‌Z.Zhen Xu,‌ S.Sergio Escalera,‌‌ A.Adrien Pavao, M.Magali Richard,‌ W.-W.Wei-Wei Tu,‌ Q.Quanming Yao,‌‌ H.Huan Zhao and I.Isabelle Guyon.‌ Codabench: Flexible, Easy-to-Use and‌ Reproducible Meta-Benchmark Platform.‌‌PatternsJuly 2022HALback to text
123‌ articleB.Burak Yelmen‌, A.Aurélien Decelle‌‌, L.Linda Ongaro, D.Davide Marnetto‌, C.Corentin Tallec‌, F.Francesco Montinaro‌‌, C.Cyril Furtlehner, L.Luca Pagani‌ and F.Flora Jay‌. Creating artificial human‌‌ genomes using generative neural networks.PLoS Genetics‌February 2021HAL DOI‌back to text
124‌‌ miscJ. K.Jared Kaplan et al..‌ Scaling Laws for Neural‌ Language Models.arXiv:2001.08361‌‌2020back to text

TAU - 2025

TAU - 2025

2025﻿​​﻿Activity reportProject-TeamTAU​​​‌

Keywords

Computer​‌﻿﻿ Science and Digital Science​​﻿﻿

Other Research Topics​​​‌ and Application Domains

1 Team​‌﻿﻿ members, visitors, external collaborators​​﻿﻿

Research Scientists

Faculty​​﻿﻿ Members

Post-Doctoral​​﻿﻿ Fellows

PhD​‌﻿﻿ Students

Technical Staff

Interns and Apprentices﻿﻿﻿‌

Administrative Assistant

Visiting﻿‌​‌ Scientist

External Collaborators﻿﻿﻿‌

2 Overall objectives​‌﻿﻿

2.1 Presentation

3 Research﻿‌​‌ program

3.1 Frugal Learning﻿​​﻿

3.2 Causal​​​‌ Learning

3.3​​﻿﻿ Machine Learning with/for Statistical​​​‌ Physics

3.4 Machine Learning for﻿​​﻿ Numerical Simulations

3.5 Challenge﻿​﻿﻿ Organization

4 Application domains​​​‌

4.1 Computational Social Sciences﻿﻿﻿‌

4.2 Energy​​​‌ Management

4.3​​​‌ Data-driven Numerical Modeling

5﻿﻿﻿‌ Social and environmental responsibility﻿‌​‌

5.1 Footprint of research﻿​​﻿ activities

5.2 Impact﻿‌​‌ of research results

6﻿﻿﻿‌ Highlights of the year﻿‌​‌

6.1 Awards

6.2﻿‌​‌ Prestigous Publications

6.3 Visibility​​​‌

7 Latest software​​﻿﻿ developments, platforms, open data​​​‌

7.1 Latest software developments﻿​﻿﻿

7.1.1 Codalab

7.1.2 Cartolabe​​﻿﻿

7.1.3 DeepHyper﻿‌​‌

7.1.4​​​‌ OmniPrint

7.1.5 codabench

7.1.6 pyriemann-qiskit

7.1.7 pyriemann

7.1.8 braindecode

7.1.9﻿‌​‌ MOABB

7.1.10 dnadna

7.2﻿‌​‌ New platforms

7.3​‌﻿﻿ Open data

8 New​​﻿﻿ results

8.1 Frugal Learning​​​‌

8.1.1​‌﻿﻿ Neural Architecture Growth for​​﻿﻿ Frugal AI

8.1.2​​​‌ Model Frugality

8.1.3 Data﻿﻿﻿‌ frugality

8.2 Toward﻿﻿﻿‌ Good AI

8.2.1 Causal﻿‌​‌ Learning

8.2.2 Explainable​‌﻿﻿ and Robust Learning

8.2.3 Improved﻿​​﻿ Learning

8.2.4 Towards​‌﻿﻿ high-quality and private genomes​​﻿﻿ based on generative neural​​​‌ networks

8.3 Machine​‌﻿﻿ Learning with/for Statistical Physics​​﻿﻿

8.3.1 Energy based models​‌﻿﻿

8.3.2 Physics Informed﻿​​﻿ Machine Learning

8.4​​​‌ Machine Learning for Numerical﻿​﻿﻿ Simulations

8.4.1 Representations and​‌﻿﻿ Reduced Order Models for​​﻿﻿ Dynamical Systems

8.4.2 Foundational﻿​​﻿ models for numerical simuations​​​‌

8.4.3 Advances in sparse﻿﻿﻿‌ recovery for inverse problem﻿‌​‌ and application in M/EEG﻿​​﻿

8.4.4 Machine​​​‌ learning for evolutionary genomics﻿​﻿﻿

8.5 Energy Management​‌﻿﻿

8.6 Computational Social​​​‌ Sciences

8.6.1 Labor Studies﻿﻿﻿‌

8.6.2​‌﻿﻿ Health and practices

8.6.3﻿‌​‌ Scientific Information System and﻿​​﻿ Visual Querying

8.7 Organization﻿​​﻿ of Challenges

8.7.1﻿‌​‌ Challenges

8.7.2​​​‌ Datasets

9 Bilateral contracts and​‌﻿﻿ grants with industry

9.1​​​‌ Bilateral contracts with industry﻿​﻿﻿

10​​​‌ Partnerships and cooperations

10.1﻿﻿﻿‌ European initiatives

10.1.1 Horizon﻿‌​‌ Europe

2025Activity reportProject-TeamTAU‌

Computer‌ Science and Digital Science

Other Research Topics‌ and Application Domains

1 Team‌ members, visitors, external collaborators

Faculty Members

Post-Doctoral Fellows

PhD‌ Students

Interns and Apprentices‌

Visiting‌‌ Scientist

External Collaborators‌

2 Overall objectives‌

3 Research‌‌ program

3.1 Frugal Learning

3.2 Causal‌ Learning

3.3 Machine Learning with/for Statistical‌ Physics

3.4 Machine Learning for Numerical Simulations

3.5 Challenge Organization

4 Application domains‌

4.1 Computational Social Sciences‌

4.2 Energy‌ Management

4.3‌ Data-driven Numerical Modeling

5‌ Social and environmental responsibility‌‌

5.1 Footprint of research activities

5.2 Impact‌‌ of research results

6‌ Highlights of the year‌‌

6.2‌‌ Prestigous Publications

6.3 Visibility‌

7 Latest software developments, platforms, open data‌

7.1 Latest software developments

7.1.2 Cartolabe

7.1.3 DeepHyper‌‌

7.1.4‌ OmniPrint

7.1.9‌‌ MOABB

7.2‌‌ New platforms

7.3‌ Open data

8 New results

8.1 Frugal Learning‌

8.1.1‌ Neural Architecture Growth for Frugal AI

8.1.2‌ Model Frugality

8.1.3 Data‌ frugality

8.2 Toward‌ Good AI

8.2.1 Causal‌‌ Learning

8.2.2 Explainable‌ and Robust Learning

8.2.3 Improved Learning

8.2.4 Towards‌ high-quality and private genomes based on generative neural‌ networks

8.3 Machine‌ Learning with/for Statistical Physics

8.3.1 Energy based models‌

8.3.2 Physics Informed Machine Learning

8.4‌ Machine Learning for Numerical Simulations

8.4.1 Representations and‌ Reduced Order Models for Dynamical Systems

8.4.2 Foundational models for numerical simuations‌

8.4.3 Advances in sparse‌ recovery for inverse problem‌‌ and application in M/EEG

8.4.4 Machine‌ learning for evolutionary genomics

8.5 Energy Management‌

8.6 Computational Social‌ Sciences

8.6.1 Labor Studies‌

8.6.2‌ Health and practices

8.6.3‌‌ Scientific Information System and Visual Querying

8.7 Organization of Challenges

8.7.1‌‌ Challenges

8.7.2‌ Datasets

9 Bilateral contracts and‌ grants with industry

9.1‌ Bilateral contracts with industry

10‌ Partnerships and cooperations

10.1‌ European initiatives

10.1.1 Horizon‌‌ Europe

10.2 National initiatives

10.2.2 Others

11.1 Promoting scientific activities‌‌

11.1.1 Scientific events: organisation

Chair of conference program‌‌ committees

Reviewer‌

11.1.2‌ Journal

Member of the‌ editorial boards

11.1.4 Leadership within the scientific‌ community

11.1.5 Scientific expertise‌

11.1.6 Research administration

11.2 Teaching‌‌ - Supervision - Juries - Educational and pedagogical‌ outreach

11.2.3 Juries‌

12 Scientific production