2025Activity report​​​‌Project-TeamMALT

4.1 Industry‌ & Telecommunication

• Heterogeneous tabular‌​‌ data generation with deep​​ generative models. Tabular data​​​‌ generation is paramount when‌ dealing with privacy-sensitive data‌​‌ and with missing values,​​ which are frequent cases​​​‌ in the real (industrial)‌ world and particularly at‌​‌ Orange. It is also​​ used for data augmentation,​​​‌ a pre-processing step often‌ needed when training data-hungry‌​‌ deep learning models (for​​ example to detect anomalies​​​‌ in networks, study customer‌ profiles, ...). The CIFRE‌​‌ PhD of G. Charbel​​ Kindji, funded by Orange,​​​‌ is concerned with this‌ application. We study methods‌​‌ to tackle this problem​​ when the tabular data​​​‌ are heterogeneous (numerical and‌ symbolic) and when new‌​‌ tables should be generated​​ from scratch based on​​​‌ a human prompt.
• Counterfactual‌ explanations over multivariate time‌​‌ series. Very complex machine​​ learning models (that are​​​‌ called black-boxes) are often‌ used in critical applications‌​‌ (e.g. self-driving cars). To​​ comply with EU regulations​​​‌ and better understand their‌ systems, many companies, and‌​‌ in particular Stellantis, are​​ interested in developing skills​​​‌ in "explainable AI", a‌ domain which aims at‌​‌ bringing back the human​​ in the decision loop​​​‌ that involves a black‌ box model. The CIFRE‌​‌ PhD of Paul Sevellec,​​ funded by Stellantis, is​​​‌ concerned with this application.‌ We study the particular‌​‌ case of counterfactual explanations​​ on the challenging context​​​‌ of multivariate time-series. This‌ problem is related to‌​‌ the generation of new​​ data that fulfills some​​​‌ human requirements.
• Analysis and‌ optimization of 3D-printing files‌​‌ through Machine Learning. In​​ the realm of Additive​​​‌ Manufacturing, and more specifically‌ Fused Filament Fabrication 3D‌​‌ printing, print time estimation​​ and optimization plays a​​​‌ pivotal role. The two‌ main approaches for this‌​‌ task are parametric models​​ taking as input the​​​‌ 3D description of the‌ model, and analytical models‌​‌ based on G-code. In​​ the context of the​​​‌ PhD of Niels Cobat‌ , we explore the‌​‌ potential of Machine Learning​​ models dedicated to sequences​​​‌ to handle this task.‌
• Transfer learning for time‌​‌ series analysis. In many​​ industrial and scientific domains—such​​​‌ as sensor monitoring, anomaly‌ detection, forecasting or activity‌​‌ recognition—time series data are​​ abundant, while labeled data​​​‌ for each new task‌ or domain remain scarce‌​‌ and costly to obtain.​​ The PhD project of​​​‌ Ambroise Odonnat addresses this‌ challenge by studying and‌​‌ developing transfer learning methods​​ specifically tailored to time​​​‌ series data, with a‌ particular focus on modern‌​‌ sequence models such as​​ Transformers. The work aims​​​‌ to design robust and‌ transferable representations that capture‌​‌ both temporal dependencies and​​ dynamic patterns, enabling knowledge​​​‌ learned from one set‌ of time series to‌​‌ be efficiently reused across​​​‌ different domains, conditions or​ applications. By bridging theoretical​‌ analysis and practical methodologies,​​ the project targets scalable​​​‌ and data-efficient learning solutions​ for real-world time series​‌ applications.
• Convergence of graph​​ and vector approaches for​​​‌ integrating machine learning predictions​ into network digital twin​‌ simulations The PhD thesis​​ of Oussama Zaid —in​​​‌ collaboration with Orange—focuses on​ the convergence of graph-based​‌ and vector-based approaches to​​ natively integrate machine learning​​​‌ predictions into digital twin​ simulations of telecommunications networks,​‌ within the context of​​ Orange’s digital twin platforms​​​‌ such as Thing’in and​ network management systems for​‌ 5G and fiber infrastructures.​​ The research aims to​​​‌ leverage graph databases and​ graph machine learning—particularly Graph​‌ Neural Networks—to model complex,​​ dynamic network structures, predict​​​‌ traffic and failures, and​ enable advanced “what-if” simulations​‌ to assess the impact​​ of network changes on​​​‌ performance and customer experience.​ Addressing current limitations where​‌ graph learning pipelines are​​ decoupled from database systems,​​​‌ the thesis proposes a​ hybrid graph–vector database that​‌ embeds machine learning capabilities​​ directly into data management.​​​‌ The expected outcome is​ a prototype supporting multi-model​‌ data representation (graphs and​​ embeddings), a unified query​​​‌ language operating across structural​ and vector spaces, and​‌ scalable storage with versioning​​ of data, models, and​​​‌ embeddings, thereby ensuring efficient​ querying, data governance, incremental​‌ learning, and reproducibility over​​ time.

4.2 Embedded Systems​​​‌

Elisa Fromont is the​ local coordinator of the​‌ ADAPTING ("Architectures adaptatives pour​​ l’intelligence artificielle embarquée") project​​​‌ of the PEPR AI.​ The ADAPTING project aims​‌ at proposing new architectural​​ paradigms adaptable to any​​​‌ AI application and its​ constraints in terms of​‌ precision, energy, latency and​​ reliability. The adaptive architecture​​​‌ will be designed to​ ensure the flexibility, efficiency,​‌ durability and reliability of​​ embedded AI. Within this​​​‌ project, she co-supervises two​ PhD students and one​‌ research apprentice (together with​​ Mathieu Lefort ) who​​​‌ respectively started in fall​ 2024 and 2025 and​‌ 2024. Within this project,​​ Simon Malinowski also co-supervises​​​‌ the PhD of Julianne​ Guerbette with the LACODAM​‌ team.

• Continual learning for​​ time series forecasting in​​​‌ embedded systems. During her​ first PhD year, Nouha​‌ Karaouli evaluated whether Time​​ Series Foundation Models (TSFMs)​​​‌ would be better suited​ for continual learning than​‌ smaller specific time series​​ models and whether they​​​‌ could meet the resource​ constraints of embedded systems.​‌ She did a systematic​​ study of the "catastrophic​​​‌ forgetting" phenomenon in TSFMs​ sequentially fine-tuned (FT) on​‌ multiple forecasting tasks, evaluating​​ TimesFM-2.0, Chronos-2, and Granite-FlowState-r1​​​‌ across synthetic and real-world​ datasets. While FT improves​‌ performance on new tasks,​​ it often degrades accuracy​​​‌ on earlier ones. All​ models exhibit forgetting, though​‌ larger models are markedly​​ more robust than smaller​​​‌ ones. Forgetting and adaptability​ further depend on the​‌ gap between pretraining and​​ fine-tuning data distributions. These​​​‌ results expose key limitations​ in current TSFM design​‌ and emphasize the need​​ for continual-learning methods for​​​‌ deployment in non-stationary forecasting​ settings.
• Trustworthy AI hardware​‌ architectures. The goal of​​ Noam Birès 's PhD​​​‌ thesis is to study​ the impact of hardware​‌ faults not only on​​ the AI decisions, but​​ also on algorithms developed​​​‌ to explain AI (XAI)‌ models. The objective is‌​‌ to make AI-HW reliable​​ by understanding how hardware​​​‌ faults (due to variability,‌ aging, external perturbations) can‌​‌ impact AI and XAI​​ decisions and how to​​​‌ mitigate those impacts efficiently.‌ The final goal is‌​‌ to enable the transparency​​ of the AI-HW by​​​‌ designing self-explainable, trustworthy, reliable,‌ and real-time verifiable AI‌​‌ hardware accelerators, capable of​​ performing self-test, self-diagnosis, and​​​‌ self-correction.
• Monitoring Federated Systems‌ with XAI.Loane Portier‌​‌ is an apprentice in​​ the AI Master of​​​‌ University of Rennes. Federated‌ Learning (FL) enables collaborative‌​‌ model training across decentralized​​ data sources, but heterogeneous​​​‌ client distributions can lead‌ to local drifts or‌​‌ adversarial behaviors that are​​ difficult to detect with​​​‌ conventional metrics. In his‌ first apprentice year (M1),‌​‌ he studied a method​​ that leverages SHAP (SHapley​​​‌ Additive exPlanations) to monitor‌ the evolution of feature‌​‌ contributions in both local​​ and global models during​​​‌ training.

4.3 Defense

• Object‌ detection from few multispectral‌​‌ examples. This project, developed​​ during the thesis of​​​‌ Manuel Nkegoum, aims at‌ providing robust deep-learning-based methods‌​‌ to detect objects in​​ outdoor environments using multispectral​​​‌ images under a low‌ supervision context. The developed‌​‌ methods are expected to​​ learn from few labeled​​​‌ examples at training time‌ and be able to‌​‌ detect scarcely-observed objects in​​ prediction. In case of​​​‌ very few object labels‌ (even no label) being‌​‌ available, the model to​​ be developed should be​​​‌ capable of discovering unknown‌ novel objects from the‌​‌ observed scene.
• Local search​​ for multi-armed bandit problems.​​​‌ Multi-armed bandits is the‌ paradigm to design algorithms‌​‌ which simultaneously learn from​​ the data they have​​​‌ collected and act (and‌ therefore collect data) based‌​‌ on what they have​​ learned. While being important​​​‌ for many aplications, such‌ algorithms prove to be‌​‌ inefficient when confronted to​​ combinatorial optimization problems. To​​​‌ remove this limit, we‌ are developing bandit algorithms‌​‌ dedicated to combinatorial problems​​ which can be solved​​​‌ through local search. This‌ project is currently supported‌​‌ by a funding from​​ a collaboration between Inria​​​‌ and DGA-AID to foster‌ reasearch subjects which are‌​‌ of interest to both​​ the army and the​​​‌ industry.

4.4 Agriculture

• Deep‌ learning-based analysis of the‌​‌ early development of bovine​​ embryos from videomicroscopy. The​​​‌ PhD of Yasmine Hachani‌ (collaboration with the Sairpico‌​‌ team and INRAE) focuses​​ on designing deep learning​​​‌ methods for the comparison‌ and classification of videos‌​‌ of embryos produced in​​ vitro (PIV). These automatic​​​‌ methods are eagerly awaited‌ by biologists in order‌​‌ to broaden the potential​​ of fundamental and applied​​​‌ research in this field,‌ and to help improve‌​‌ results and reproductive performance​​ in breeding. The problem​​​‌ posed is multifaceted. First‌ of all, the images‌​‌ acquired by microscopy are​​ complex in nature: they​​​‌ are low-contrast, noisy, contain‌ transparency effects, and movements‌​‌ are difficult to characterize.​​ The categorization of in​​​‌ vitro fertilized embryos, in‌ terms of the quality‌​‌ of their development, is​​ based on a continuum​​​‌ of classes, rather than‌ distinct ones. Furthermore, the‌​‌ need is to obtain​​​‌ reliable classification at the​ earliest possible stage, i.e.​‌ 3 days post-gamete contact,​​ from a video of​​​‌ 300 images, with images​ acquired every 15 minutes.​‌ Finally, while classification can​​ be supervised, we have​​​‌ only a limited amount​ of data (a few​‌ hundred videos) for deep​​ learning purposes, especially as​​​‌ class characterization can only​ be achieved by observing​‌ a video in its​​ entirety.

4.5 Medicine

• On​​​‌ multimodal segmentation of chronic​ stroke lesions. Since fall​‌ 2024, Elisa Fromont is​​ the co-supervisor of Youwan​​​‌ Mahé with the EMPENN​ Inria team. Post-stroke MRI​‌ not only delineates focal​​ lesions but also reveals​​​‌ secondary structural changes, such​ as atrophy and ventricular​‌ enlargement. These abnormalities, increasingly​​ recognised as imaging biomarkers​​​‌ of recovery and outcome,​ remain poorly captured by​‌ supervised segmentation methods. In​​ this thesis, we adapt​​​‌ state-of-the-art generative methods to​ this particular medical context​‌ to better detect and​​ segment these anomalies.

5.1 Footprint of research​‌ activities

There are two​​ main axes that characterize​​​‌ the bulk of MALT's​ environmental impact: work trips,​‌ and computing resources utilisation.​​

• Work trips. Whenever possible,​​​‌ we prefer using train​ rather than plane for​‌ national and European travels.​​ Most of us continue​​​‌ to submit papers to​ international conferences outside of​‌ Europe but if a​​ paper gets accepted into​​​‌ such conference, we prioritize​ sending the first author​‌ (PhD student). Outside of​​ conferences, for national events​​​‌ (seminars, PhD juries, etc.),​ videoconference is increasingly used,​‌ which helps to reduce​​ the overall carbon footprint​​​‌ of the community.
• Utilisation​ of computing resources. The​‌ discontinuation of Igrida services​​ and the transition towards​​​‌ Grid'5000 and Jean Zay​ has reduced our access​‌ to easily available computation​​ resources. It adds friction​​​‌ to making experiments, but​ has a positive effect​‌ on energy consumption, as​​ we are now using​​​‌ national infrastructures that benefit​ from even better sharing​‌ between users than Igrida​​ (which was already heavily​​​‌ used).

5.2 Impact of​ research results

We estimate​‌ that the research work​​ can have actual impact​​​‌ in three different ways:​

• In the short/medium term,​‌ a significant part of​​ our research work is​​​‌ conducted in collaboration with​ companies, through CIFRE PhDs.​‌ Hence, the addressed research​​ problems concern an important​​​‌ challenge for the company,​ and the solutions proposed​‌ are evaluated on their​​ relevance to tackle this​​​‌ challenge.
• In the long​ term, the team has​‌ a fundamental line of​​ work on trustworthy and​​​‌ interpretable machine learning. Given​ the increasing use of​‌ machine learning solutions in​​ most areas of human​​​‌ activity, work on interpretability​ is of utmost societal​‌ importance, as it will​​ help in designing more​​​‌ useful and also more​ acceptable machine learning approaches.​‌ This will require a​​ sustained effort from the​​​‌ community: MALT is taking​ part in this effort​‌ with an important number​​ of contributions in this​​​‌ area.

7.1‌ Latest software developments

8.1 Learn on​​ temporal data

Participants: Patrick​​​‌ Bouthemy, Élisa Fromont‌, Romaric Gaudel,‌​‌ Nouha Karaouli, Paul​​ Sévellec, Yasmine Hachani​​​‌.

Remark about the‌ “Participants” boxes: we compiled‌​‌ syntactically the list of​​ co-authors of the papers​​​‌ that make the “New‌ Results” of the year,‌​‌ for each section. It​​ obviously does not mean​​​‌ that other members of‌ the team do not‌​‌ work on the topics​​ listed, the correct meaning​​​‌ is that they did‌ not have a publication‌​‌ on that topic this​​ year.

During the year,​​​‌ we have proposed several‌ advancements in temporal learning.‌​‌ In time series forecasting,​​ we demonstrated that foundation​​​‌ models exhibit inherent limitations‌ for time series data‌​‌ due to domain-dependent zero-shot​​ capabilities, revealing that fine-tuned​​​‌ models often fail to‌ deliver consistent performance gains‌​‌ over dedicated architectures despite​​ their larger parameter counts​​​‌ (see 25). For‌ multivariate time series classification,‌​‌ we introduced CFE4MTS,​​ a generation-based counterfactual explanation​​​‌ framework that produces interpretable,‌ plausible counterfactuals aligned with‌​‌ target class distributions, significantly​​ outperforming state-of-the-art methods across​​​‌ real-world datasets (see 23‌). Additionally, we uncovered‌​‌ critical privacy risks in​​ fine-grained electrical consumption data,​​​‌ showing that even degraded‌ time series (e.g., rounded‌​‌ to 100W) can re-identify​​ over 40% of households​​​‌ using just 7 consecutive‌ measurements, with uniqueness rates‌​‌ persisting at concerning levels​​ even under severe data​​​‌ degradation (see 12).‌ In videomicroscopy, we introduced‌​‌ CLEmbryo to identify the​​ developmental stages of embryos​​​‌ from 2D time-lapse image‌ sequences. This model leverages‌​‌ supervised contrastive learning with​​ focal loss and the​​​‌ lightweight 3D neural network‌ CSN-50 as encoder (see‌​‌ 29).

8.2 Learn​​ through time

Participants: Ambroise​​​‌ Odonnat, Romain Tavenard‌.

During the year,‌​‌ we advanced the "Learn​​ through time" initiative by​​​‌ introducing four key contributions.‌ First, we proposed Zero-shot‌​‌ Model-based Reinforcement Learning using​​ Large Language Models15​​​‌ to enable LLMs in‌ predicting continuous dynamics via‌​‌ Disentangled In-Context Learning (DICL).​​ Second, we developed PAWL​​​‌17, an efficient‌ algorithm for exact partial‌​‌ Wasserstein distances on the​​ line, and its differentiable​​​‌ extension for high-dimensional settings‌ 24. Third, we‌​‌ released SKADA-Bench10,​​​‌ a comprehensive benchmark for​ evaluating unsupervised domain adaptation​‌ across diverse modalities. Finally,​​ we introduced Leveraging Gradients​​​‌ for Unsupervised Accuracy Estimation​ under Distribution Shift13​‌, a technique that​​ uses gradient magnitudes to​​​‌ predict test accuracy under​ distribution shifts and significantly​‌ outperforms state-of-the-art methods.

8.3​​ Trustworthy AI

Participants: Élisa​​​‌ Fromont, Romaric Gaudel​, G. Charbel Kindji​‌, Dimitri Lerévérend,​​ Paul Sévellec, Paul​​​‌ Viallard.

During the​ year, we have proposed​‌ Zip-DL, a privacy-aware decentralized​​ learning algorithm that achieves​​​‌ $\mathbf{35}\%$ higher membership-inference​ attack resilience compared to​‌ baseline methods while maintaining​​ up to $\mathbf{59}\%‌$ higher accuracy under the​ same threat model, as​‌ demonstrated by 16.​​ Concurrently, to address analytical​​​‌ variability in fMRI processing,​ we released the HCP​‌ Multi-Pipeline dataset comprising $\mathbf{1},\mathbf{080}$ participants' statistic​​​‌ maps across $\mathbf{24}$ pipelines​ and developed a style​‌ transfer framework using GANs​​ and diffusion models to​​​‌ enhance reproducibility, as detailed​ in 8 and 18​‌. In the realm​​ of data integrity, we​​​‌ advanced synthetic tabular data​ detection by introducing cross-table​‌ baseline detectors and robust​​ schema-adaptive methods 26,​​​‌ demonstrating the challenge of​ $\mathrm{𝑖𝑛}-\mathrm{𝑡ℎ𝑒}-‌\mathrm{𝑤𝑖𝑙𝑑}$ detection across diverse​​ table structures. Additionally, we​​​‌ established a PAC-Bayesian framework​ linking flat minima to​‌ generalization, showing their positive​​ influence via gradient-based bounds​​​‌ that avoid dimension-dependent dependencies​ 19; extended counterfactual​‌ explanations for multivariate time​​ series classification through plausible​​​‌ generation techniques, 23;​ and quantified privacy risks​‌ in fine-grained electrical consumption​​ data, revealing that $\mathbf{90‌}\%$ of households can​ be re-identified from $\mathbf{5‌}$ consecutive measurements—a critical insight​​ for smart grid privacy—​​​‌12.

8.4 Applications​

Participants: Patrick Bouthemy,​‌ Élisa Fromont, Romaric​​ Gaudel, G. Charbel​​​‌ Kindji, Youwan Mahé​, Ambroise Odonnat,​‌ Paul Sévellec, Simon​​ Malinowski, Erwan Vincent​​​‌.

During the year,​ we have proposed significant​‌ advancements across medical imaging,​​ tabular data processing, and​​​‌ foundational machine learning techniques.​ In medical imaging, we​‌ developed a deep learning​​ framework for multi-modal MRI​​​‌ segmentation of sub-acute and​ chronic stroke lesions, achieving​‌ a mean Dice score​​ of 0.74 for dual-modality​​​‌ (T1-w + FLAIR) ensembles​ on internal datasets—demonstrating that​‌ integrating FLAIR data and​​ ensemble strategies significantly improves​​​‌ small/medium lesion quantification 7​. Concurrently, we introduced​‌ the HCP Multi-Pipeline dataset​​ to investigate analytical variability​​​‌ in fMRI, providing 1,080​ participants' statistic maps across​‌ 24 pipelines for rigorous​​ analysis of processing inconsistencies​​​‌ 8. To address​ reproducibility challenges, we further​‌ proposed a style transfer​​ approach leveraging GANs and​​​‌ Diffusion Models to convert​ fMRI statistic maps across​‌ pipelines, enabling effective data​​ augmentation for neuroimaging studies​​​‌ 18. In tabular​ data processing, we conducted​‌ the first comprehensive benchmark​​ of 16 diverse datasets,​​​‌ revealing that diffusion-based generative​ models consistently outperform alternatives​‌ after dataset-specific tuning—while highlighting​​ the challenges of cross-table​​​‌ synthetic data detection under​ schema variability 9,​‌ 26, 20.​​ Additionally, we developed the​​​‌ supervised model CLEmbryo for​ cell stage classification of​‌ animal embryos in videomicroscopy,​​ which outperforms state-of-the-art methods​​ on both our in-house​​​‌ Bovine ECS dataset and‌ the publicly available NYU‌​‌ Mouse Embryos dataset 29​​; extended our understanding​​​‌ of optimization pathways through‌ a circuit-based curriculum for‌​‌ efficient learning 22;​​ and developed plausible counterfactual​​​‌ explanations for multivariate time‌ series classification 23.‌​‌ Finally, we demonstrated that​​ gradient-based signal analysis (via​​​‌ cross-entropy loss gradients) provides‌ robust test accuracy estimation‌​‌ under distribution shift, outperforming​​ existing methods in diverse​​​‌ scenarios 13. These‌ contributions collectively advance robustness‌​‌ in medical data, scalable​​ tabular generation, and interpretable​​​‌ AI systems for real-world‌ deployment. In 14,‌​‌ we have developed a​​ methodology that is able​​​‌ to identify impact factors‌ for buses commercial speed‌​‌ analysis. Machine learning and​​ data analysis techniques are​​​‌ used to quantify the‌ impact of many different‌​‌ features on the commercial​​ speed of buses. The​​​‌ most important features can‌ then be used within‌​‌ machine learning frameworks in​​ order to predict the​​​‌ commercial speed of buses‌ on different roads.

9.1 Bilateral​​​‌ contracts with industry

• ORANGE‌ - Univ. Rennes (2023-2025)‌​‌

  Participants: Elisa Fromont,​​ G. Charbel Kindji.​​​‌

  Contract amount: 45k€‌ + Phd Salary

  Context.‌​‌ Tabular data generation is​​ paramount when dealing with​​​‌ privacy-sensitive data and with‌ missing values, which are‌​‌ frequent cases in the​​ real (industrial) world and​​​‌ particularly at Orange. It‌ is also used for‌​‌ data augmentation, a pre-processing​​ step often needed when​​​‌ training data-hungry deep learning‌ models (for example to‌​‌ detect anomalies in networks,​​ study customer profiles, ...).​​​‌

  Objective. We study methods‌ to tackle heterogeneous tabular‌​‌ data generation with deep​​ generative models. We are​​​‌ particularly interested in problems‌ where the tabular data‌​‌ are heterogeneous (numerical and​​ symbolic) and when new​​​‌ tables should be generated‌ from scratch based on‌​‌ a human prompt.

  Additional​​ remarks. This is the​​​‌ doctoral contract for the‌ PhD of Charbel Kindji‌​‌ who defended his PhD​​ December 18 2025 (Thèse​​​‌ CIFRE).

• Stellantis - Univ.‌ Rennes (2024-2026) with LACODAM‌​‌ Team

  Participants: Elisa Fromont​​, Romaric Gaudel,​​​‌ Paul Sevellec.

  Contract‌ amount: 70k€ +‌​‌ Phd Salary

  Context.​​ This project is a​​​‌ collaboration with Stellantis and‌ focuses on the development‌​‌ of interpretable machine learning​​ models for multivariate time​​​‌ series data. Utilizing a‌ range of sensors integrated‌​‌ within vehicles, these models​​ are designed to make​​​‌ real-time decisions. Providing drivers‌ with clear explanations of‌​‌ these decisions is a​​ key aspect. We specifically​​​‌ concentrate on counterfactual explanations,‌ which not only clarify‌​‌ why a particular decision​​ was made but also​​​‌ illustrate how alternative scenarios‌ might have led to‌​‌ different outcomes.

  Objective. Current​​ approaches providing counterfactual explanations​​​‌ for time series models‌ are limited to univariate‌​‌ time series. In this​​ project, we aim to​​​‌ develop approaches to handle‌ multivariate time series, which‌​‌ requires capturing the correlations​​ between the series.

  Additional​​​‌ remarks. This is the‌ doctoral contract for the‌​‌ PhD of Paul Sévellec​​ (Thèse CIFRE), which is​​​‌ co-advised with Laurence Rozé‌ from LACODAM Team.

• ATERMES‌​‌ - Univ. Rennes (2024-2027)​​​‌ with OBELIX Team

  Participants:​ Élisa Fromont, Manuel​‌ Nkegoum.

  Contract amount​​: 0€ (for MALT​​​‌ Team) + Phd Salary​

  Objective. This project​‌ aims at providing robust​​ deep-learning-based methods to detect​​​‌ objects in outdoor environments​ using multispectral images under​‌ a low supervision context.​​ The developed methods are​​​‌ expected to learn from​ few labeled examples at​‌ training time and be​​ able to detect scarcely-observed​​​‌ objects in prediction. In​ case of very few​‌ object labels (even no​​ label) being available, the​​​‌ model to be developed​ should be capable of​‌ discovering unknown novel objects​​ from the observed scene.​​​‌

  Additional remarks. This is​ the CIFRE PhD of​‌ Manuel Nkegoum with Atermes​​ (Thèse CIFRE). There is​​​‌ an agreement with the​ Obelix team to freely​‌ use part of the​​ 60k€ contract as was​​​‌ done conversely in the​ previous PhD with the​‌ same parties.

• SIEMENS -​​ Univ. Rennes (2025-2028) with​​​‌ EMPENN Team

  Participants: Elisa​ Fromont, Youwan Mahé​‌.

  Contract amount:​​ 12k€ (for MALT Team)+​​​‌ Phd Salary

  Context.​ Stroke is a major​‌ health issue globally, causing​​ severe brain damage due​​​‌ to disrupted blood supply.​ Medical imaging, especially MRI,​‌ is crucial for assessing​​ stroke localization and extent.​​​‌

  Objective. Objective.​ Our goal in this​‌ project, is to improve​​ the detection and delineation​​​‌ of chronic stroke lesions​ from multimodal data using​‌ deep learning, helping clinicians​​ plan better treatment and​​​‌ rehabilitation programs.

  Additional remarks.​ This is the CIFRE​‌ PhD of Youwan Mahé​​ with Siemens (Thèse CIFRE).​​​‌ The total contract with​ Siemens is 50k€ but​‌ this amount is divided​​ between the CHU of​​​‌ Rennes, the Empenn team​ and the MALT team.​‌

• HUAWEI - Univ. Rennes​​ II (2025-2028) with OBELIX​​​‌ Team

  Participants: Romain Tavenard​, Ambroise Odonnat.​‌

  Contract amount: 90k€​​ (for MALT Team, shared​​​‌ with OBELIX team in​ practice)+ Phd Salary

  Objective​‌. Our goal in​​ this project is to​​​‌ understand how Transformer architectures​ are impacted by distribution​‌ shift so as to​​ be able to better​​​‌ adapt pre-trained models on​ target distributions.

  Additional remarks.​‌ This is the CIFRE​​ PhD of Ambroise Odonnat​​​‌ with Huawei (Thèse CIFRE).​ There is an agreement​‌ with the Obelix team​​ to freely share the​​​‌ 90k€ contract, though this​ contract is hosted on​‌ the Université de Rennes​​ II - MALT side.​​​‌

• ORANGE - Univ. Rennes​ (2025-2028)

  Participants: Romaric Gaudel​‌, Oussama Zaid.​​

  Contract amount: 39k€​​​‌ + Phd Salary

  Objective​. In this project​‌ we develop a hybrid​​ graph–vector database that natively​​​‌ integrates machine learning, particularly​ graph neural networks. Its​‌ main use-case is telecom​​ network digital twins, aiming​​​‌ at efficient prediction, simulation,​ and data governance for​‌ complex and evolving networks.​​

  Additional remarks. This is​​​‌ the CIFRE PhD of​ Oussama Zaid with Orange​‌ (Thèse CIFRE).

10.1 International​​​‌ initiatives

10.2 International research visitors‌​‌

10.3 National​​​‌ initiatives

• DGA-AID : Local‌ search for multi-armed bandit‌​‌ problems - Inria

  Participants:​​ Romaric Gaudel, Elisa​​​‌ Fromont, Paul Viallard‌, Aurélien Delage.‌​‌

  Budget: 130k€ (Inria)

  This​​ project aims at proposing​​​‌ multi-armed bandit algorithms dedicated‌ to combinatorial problems which‌​‌ can be solved through​​​‌ local search. It is​ funding by through a​‌ collaboration between Inria and​​ DGA-AID to foster reasearch​​​‌ subjects which interest either​ the army or the​‌ industry. The fund mainly​​ covers a 2-years postdoc​​​‌ position.

• PEPR IA ADAPTING​ - Univ. Rennes (2024-2028)​‌

  Participants: Elisa Fromont,​​ Nouha Karaouli, Noam​​​‌ Bires, Loane Portier​, Simon Malinowski,​‌ Julianne Guerbette.

  Budget:​​ 3$\times$(14,5k€ +​​​‌ PhD Salary)

  AdaptING explores​ new models, computing paradigms​‌ (i.e., beyond the Von​​ Neumann architecture), hybrid architectures​​​‌ (i.e., beyond MPSoC –​ System-on-Chip), and emerging technologies​‌ through various initiatives aimed​​ at making AI more​​​‌ efficient, sustainable, and trustworthy.​ While the project encompasses​‌ hardware advancements, our contributions​​ in LACODAM will focus​​​‌ on the algorithmic level.​ In particular, we will​‌ design new resource-efficient incremental​​ learning algorithms that can​​​‌ run on embedded systems​ with their associated resource​‌ and privacy constraints. We​​ will also investigate post-hoc​​​‌ explanation methods for federated​ learning systems as a​‌ way to monitor the​​ trustworthiness of such systems.​​​‌ Federated learning will often​ be at the center​‌ of the project as​​ a practical learning paradigm​​​‌ suited for embedded systems.​

  We currently supervise three​‌ PhD students funded by​​ the PEPR IA ADAPTING.​​​‌

• ANR DATES - Univ.​ Rennes (2025-2029)

  Participants: Romain​‌ Tavenard, Romaric Gaudel​​, Elisa Fromont.​​​‌

  Budget: 238k€

  DATeS focuses​ on developing data-efficient machine​‌ learning methods for time-dependent​​ data that can generalize​​​‌ across varying conditions, environments,​ or application domains. Its​‌ application context lies in​​ real-world settings where temporal​​​‌ data are continuously collected—such​ as industrial systems, monitoring​‌ infrastructures, or complex cyber-physical​​ systems—but where annotated data​​​‌ are scarce or heterogeneous.​ The project addresses the​‌ challenge of distribution shifts​​ and domain changes by​​​‌ designing learning frameworks that​ can adapt knowledge learned​‌ from one context to​​ another. By combining methodological​​​‌ advances with practical use​ cases, it aims to​‌ improve the robustness, reliability,​​ and scalability of time-series–based​​​‌ decision systems. Overall, the​ project targets impactful applications​‌ where adaptive and transferable​​ temporal models can significantly​​​‌ reduce deployment costs and​ improve operational performance. DATeS​‌ is a joint PRCE​​ project with Univ. Jean​​​‌ Monnet (St Etienne) and​ ERICSSON R&D.

• P16 -​‌ Inria (2025-2026)

  Participants: Romain​​ Tavenard.

  Budget: Engineer​​​‌ Salary

  The tslearn library​ that is developped in​‌ the team is backed​​ by a P16 Inria​​​‌ project through the funding​ of the engineer contract​‌ of Guillaume Charavel.

• AEx​​ HYPE: HYPErparameter-Free Optimization Algorithms​​​‌ by Online Self-Tuning -​ Inria (2025-2028)

  Participants: Paul​‌ Viallard, Romaric Gaudel​​.

  Budget: 146k€ (Inria)​​​‌

  This project lies at​ the intersection of statistical​‌ learning theory and mathematical​​ optimization, both central topics​​​‌ in machine learning. Indeed,​ optimization algorithms are the​‌ backbone of machine learning​​ methods, enabling us to​​​‌ automatically find a model​ (i.e. a mathematical function)​‌ from data to perform​​ a desired task. However,​​​‌ running these optimization algorithms​ requires setting certain hyperparameters​‌ that influence their execution,​​ and identifying optimal values​​​‌ for them can be​ time-consuming. Therefore, the goal​‌ of this project is​​ to develop novel optimization​​ algorithms capable of adaptively​​​‌ tuning all their hyperparameters‌ during execution. The funding‌​‌ provides support for a​​ three-year PhD position.

10.4​​​‌ Regional initiatives

• AIS fund‌ for 3D-printing time prediction‌​‌

  Participants: Romaric Gaudel.​​

  Budget: 25k€ (Rennes Métropole:​​​‌ Allocation d'Insatallation Scientifique)

  This‌ fund pays for a‌​‌ GPU card used for​​ research on improving 3D​​​‌ printing. The 3D printing‌ research axis emerged within‌​‌ the IRISA laboratory in​​ 2021, led by Damien​​​‌ Hardy and Fabrice Lamarche,‌ and is still in‌​‌ the process of consolidation.​​ As part of this​​​‌ emerging research axis, Damien‌ Hardy, Romaric Gaudel ,‌​‌ and Niels Cobat work​​ at optimizing G-code using​​​‌ AI methods, in particular‌ neural networks for sequence‌​‌ modeling, and this research​​ track benefits from the​​​‌ funded GPU.

  Romaric AIS‌

11.1 Promoting‌​‌ scientific activities

11.2​‌ Teaching - Supervision -​​ Juries - Educational and​​​‌ pedagogical outreach

11.3 Popularization

12.1 Major publications​​

• 1 inproceedingsS.Sayan​​​‌ Biswas, D.Davide‌ Frey, R.Romaric‌​‌ Gaudel, A.-M.Anne-Marie​​ Kermarrec, D.Dimitri​​​‌ Lerévérend, R.Rafael‌ Pires, R.Rishi‌​‌ Sharma and F.François​​ Taïani. Low-Cost Privacy-Preserving​​​‌ Decentralized Learning.Proceedings‌ on Privacy Enhancing Technologies‌​‌ SymposiumPrivacy Enhancing Technologies​​ Symposium20253Washinghton​​​‌ DC, United StatesJuly‌ 2025, 451–474HAL‌​‌DOI
• 2 inproceedingsL.​​Laetitia Chapel and R.​​​‌Romain Tavenard. One‌ for all and all‌​‌ for one: Efficient Computation​​ of Partial Wasserstein Distances​​​‌ on the Line.‌Proceedings of the International‌​‌ Conference on Learning Representations​​International Conference on Learning​​​‌ RepresentationsSingapore, SingaporeApril‌ 2025HAL
• 3 inproceedings‌​‌L.Laetitia Chapel,​​ R.Romain Tavenard and​​​‌ S.Samuel Vaiter.‌ Differentiable Generalized Sliced Wasserstein‌​‌ Plans.NeurIPS 2025​​ - 39th Annual Conference​​​‌ on Neural Information Processing‌ SystemsSan Diego, United‌​‌ States2025, 1-20​​HALDOI
• 4 article​​​‌G.G.Charbel Kindji,‌ L. M.Lina Maria‌​‌ Rojas-Barahona, E.Elisa​​ Fromont and T.Tanguy​​​‌ Urvoy. Tabular Data‌ Generation Models: An In-Depth‌​‌ Survey and Performance Benchmarks​​ with Extensive Tuning.​​​‌Neurocomputing658December 2025‌, 131655HALDOI‌​‌
• 5 inproceedingsP.Paul​​ Sevellec, E.Elisa​​​‌ Fromont, R.Romaric‌ Gaudel, L.Laurence‌​‌ Rozé and M.Matteo​​ Sammarco. Plausible Conditional​​​‌ Generation-based Counterfactual Explanations for‌ Multivariate Times Series Classification‌​‌.Proceedings of the​​ 28th European Conference on​​​‌ Artificial Intelligence, IOS Press‌ECAI 2025 - European‌​‌ Conference on Artificial Intelligence​​Bologna, ItalyOctober 2025​​​‌HAL
• 6 articleA.‌Antonin Voyez, T.‌​‌Tristan Allard, G.​​Gildas Avoine, P.​​​‌Pierre Cauchois, E.‌Elisa Fromont and M.‌​‌Matthieu Simonin. The​​​‌ Privacy Cost of Fine-Grained​ Electrical Consumption Data.​‌Scientific Reports1517391​​May 2025, 1-8​​​‌HALDOI

12.2 Publications​ of the year