2025​‌Activity reportProject-TeamTHOTH​​

Compute​‌

A significant amount of​​ the team’s computations are​​​‌ performed on Jean Zay​ national cluster. According to​‌ the cluster's reporting platform,​​ 50k normalized GPU hours​​​‌ have been used by​ the team, which amounts​‌ to 1.2 tons eqCO2.​​ Besides computations performed on​​​‌ this cluster, the team​ maintained its own cluster,​‌ on which part of​​ the computations are done​​​‌ as well. Assuming 10​ GPUs are used at​‌ all times (which is​​ a rather generous estimate),​​​‌ this amounts to less​ than 100k GPU hours​‌ over the year. Most​​ of these machines are​​​‌ hosted in the datacenter​ of the IMAG building,​‌ which is probably slightly​​ less efficient than the​​​‌ GENCI infrastructure. Overall, we​ estimate our local consumption​‌ to be under 3​​ tons eqCO2.

In total,​​​‌ we estimate the emissions​ of the team's compute​‌ to be about 4​​ tons eqCO2. While we​​​‌ do not provide impact​ if term of resources,​‌ the team dedicated a​​ special effort to keep​​ local computing servers running​​​‌ for as long as‌ possible, upgrading them when‌​‌ possible to avoid replacing​​ them.

This does not​​​‌ count the Dino (V2‌ and V3) models, which‌​‌ are significantly more expensive​​ to train but are​​​‌ also significantly more impactful‌ than an average research‌​‌ paper, being used in​​ more than 10K scientific​​​‌ projects, and for which‌ full emissions data is‌​‌ available.

Travel

The other​​ main CO2eq footprint is​​​‌ international flights. While we‌ did not gather specific‌​‌ numbers, team members take​​ special care in reducing​​​‌ their plane travels (several‌ permanent researchers have not‌​‌ traveled by plane for​​ several years), refusing distant​​​‌ invitations, as well as‌ encouraging less travel-hungry community‌​‌ practices. This has led​​ to a drastic reduction​​​‌ of our travel impact‌ over the years, which‌​‌ we will try to​​ quantify for the next​​​‌ activity report.

7.1.1 Cyanure‌​‌

• Name:
  Cyanure: An Open-Source​​ Toolbox for Empirical Risk​​​‌ Minimization
• Functional Description:
  Cyanure‌ is an open-source C++‌​‌ software package with a​​ Python interface. The goal​​​‌ of Arsenic is to‌ provide state-of-the-art solvers for‌​‌ learning linear models, based​​ on stochastic variance-reduced stochastic​​​‌ optimization with acceleration mechanisms‌ and Quasi-Newton principles. Arsenic‌​‌ can handle a large​​​‌ variety of loss functions​ (logistic, square, squared hinge,​‌ multinomial logistic) and regularization​​ functions (l2, l1, elastic-net,​​​‌ fused Lasso, multi-task group​ Lasso). It provides a​‌ simple Python API, which​​ is very close to​​​‌ that of scikit-learn, which​ should be extended to​‌ other languages such as​​ R or Matlab in​​​‌ a near future.
• Release​ Contributions:
  packaging on conda​‌ and pipy + various​​ improvements
• URL:
  http://­thoth.­inrialpes.­fr/­people/­mairal/­arsenic/­welcome.­html
• Contact:​​​‌
  Julien Mairal
• Participant:
  2​ anonymous participants

7.1.2 MLXP​‌

• Name:
  Machine Learning eXperimentalist​​ for Python
• Keywords:
  Reproducibility,​​​‌ Replication and consistency, Machine​ learning
• Functional Description:
  MLXP​‌ is an open-source, simple,​​ and lightweight experiment management​​​‌ tool based on Python.​ It streamlines the experimental​‌ process with minimal practitioner​​ overhead while ensuring a​​​‌ high level of reproducibility.​ As an open-source package,​‌ MLXP facilitates experiment launching,​​ logging, and efficient result​​​‌ exploitation. Key components include​ automated job launching and​‌ hierarchical configuration files, logging​​ of experiment outputs along​​​‌ with metadata, automated code​ and job version management,​‌ seamless multi-job submission to​​ a HPC job scheduler,​​​‌ and intuitive result exploitation​ capabilities including querying results,​‌ grouping and aggregation operations.​​
• URL:
  https://­inria-thoth.­github.­io/­mlxp/­pages/­master/­index.­html
• Contact:
  Michael​​​‌ Arbel

Object-wise​‌ Distance Estimation for Event​​ Camera Data

Participants: Nan​​​‌ Cai, Pia Bideau​.

Event cameras provide​‌ a natural and data​​ efficient representation of visual​​​‌ information, motivating novel computational​ strategies towards extracting visual​‌ information. Inspired by the​​ biological vision system, in​​​‌ this work 26 propose​ a behavior driven approach​‌ for object-wise distance estimation​​ from event camera data.​​​‌ This behavior-driven method mimics​ how biological systems, like​‌ the human eye, stabilize​​ their view based on​​​‌ object distance: distant objects​ require minimal compensatory rotation​‌ to stay in focus,​​ while nearby objects demand​​​‌ greater adjustments to maintain​ alignment. This adaptive strategy​‌ leverages natural stabilization behaviors​​ to estimate relative distances​​​‌ effectively. Unlike traditional vision​ algorithms that estimate depth​‌ across the entire image,​​ our approach targets local​​​‌ depth estimation within a​ specific region of interest.​‌ By aligning events within​​ a small region, we​​​‌ estimate the angular velocity​ required to stabilize the​‌ image motion. We demonstrate​​ that, under certain assumptions,​​​‌ the compensatory rotational flow​ is inversely proportional to​‌ the object's distance. The​​ proposed approach achieves new​​​‌ state-of-the-art accuracy in distance​ estimation on the dataset​‌ EVIMO2.

Figure

Salience-SGG: Enhancing​ Unbiased Scene Graph Generation​‌ with Iterative Salience Estimation​​

Participants: Runfeng Qu,​​​‌ Ole Hall, Pia​ Bideau, Julie Ouerfelli-Ethier​‌, Martin Rolfs,​​ Klaus Obermayer, Olaf​​​‌ Hellwich.

Scene Graph​ Generation (SGG) suffers from​‌ a long-tailed distribution, where​​ a few predicate classes​​​‌ dominate while many others​ are underrepresented, leading to​‌ biased models that underperform​​ on rare relations. Unbiased-SGG​​​‌ methods address this by​ implementing debiasing strategies, but​‌ often at the cost​​ of spatial understanding—resulting in​​​‌ over-reliance on semantic priors.​ In 37, we​‌ introduce Salience-SGG, a novel​​ framework featuring an Iterative​​​‌ Salience Decoder (ISD) that​ emphasizes triplets with salient​‌ spatial structures. To support​​ this, we propose semantic-agnostic​​ salience labels guiding ISD.​​​‌ Evaluations on Visual Genome,‌ Open Images V6, and‌​‌ GQA-200 show that Salience-SGG​​ achieves state-of-the-art performance and​​​‌ improves existing Unbiased-SGG methods‌ in their spatial understanding‌​‌ as demonstrated by the​​ Pairwise Localization Average Precision.​​​‌

Code is available on‌ github.

Figure

Watching Swarm Dynamics from​​ Above: A Framework for​​​‌ Advanced Object Tracking in‌ Drone Videos

Participants: Pia‌​‌ Bideau, Duc Pham​​, Félicie Dhellemmes,​​​‌ Matthew Hansen, Jens‌ Krause.

Easily accessible‌​‌ technologies, such as drones​​ equipped with diverse onboard​​​‌ sensors, have greatly expanded‌ opportunities to study animal‌​‌ behavior in natural environments.​​ However, analyzing large volumes​​​‌ of unlabeled video data,‌ often spanning hours, remains‌​‌ a significant challenge for​​ machine learning, particularly in​​​‌ computer vision. Existing approaches‌ typically process only a‌​‌ small number of frames,​​ and accurate georeferencing of​​​‌ tracked positions is still‌ largely unresolved, particularly in‌​‌ dynamic environments where static​​ landmarks cannot be established.​​​‌ In this work, we‌ focus on long-term tracking‌​‌ of animal behavior in​​ real-world geographic coordinates. To​​​‌ address this challenge, we‌ utilize classical probabilistic methods‌​‌ for state estimation, such​​ as particle filtering. Particle​​​‌ filters offer a useful‌ algorithmic structure for recursively‌​‌ adding new incoming information​​ and thus ensuring time​​​‌ consistency. By incorporating recent‌ developments in semantic object‌​‌ segmentation, we enable continuous​​ tracking of rapidly evolving​​​‌ object formations, even in‌ scenarios with limited data‌​‌ availability. We propose a​​ novel approach for tracking​​​‌ schools of fish in‌ the open ocean from‌​‌ drone videos. Our framework​​ not only performs classical​​​‌ object tracking in image‌ coordinates, instead it additionally‌​‌ tracks the position and​​ spatial expansion of the​​​‌ fish school in geographic‌ coordinates by fusing video‌​‌ data and the drone's​​ on board sensor information​​​‌ (GPS and IMU). No‌ landmarks with known geographic‌​‌ coordinates are required, making​​ the proposed method adaptable​​​‌ to unstructured, dynamic environments‌ like the open ocean,‌​‌ where static landmarks are​​ unavailable. With this, the​​​‌ presented framework enables researchers‌ to study the collective‌​‌ behavior of fish schools​​ within their social and​​​‌ environmental context.

Code and‌ the newly introduced dataset‌​‌ for tracking collective animal​​ behavior over long time​​​‌ horizons in marine environments‌ are available here.‌​‌

Figure

LUDVIG:​​ Learning-free Uplifting of 2D​​​‌ Visual features to Gaussian‌ Splatting scenes.

Participants: Juliette‌​‌ Marrie, Romain Menegaux​​, Michael Arbel,​​​‌ Diane Larlus, Julien‌ Mairal.

In 34‌​‌, we address the​​ problem of extending the​​​‌ capabilities of vision foundation‌ models such as DINO,‌​‌ SAM, and CLIP, to​​ 3D tasks. Specifically, we​​​‌ introduce a novel method‌ to uplift 2D image‌​‌ features into 3D Gaussian​​ Splatting scenes. Unlike traditional​​​‌ approaches that rely on‌ minimizing a reconstruction loss,‌​‌ our method employs a​​​‌ simpler and more efficient​ feature aggregation technique, augmented​‌ by a graph diffusion​​ mechanism. Graph diffusion enriches​​​‌ features from a given​ model, such as CLIP,​‌ by leveraging 3D geometry​​ and pairwise similarities induced​​​‌ by another strong model​ such as DINOv2. Our​‌ approach achieves performance comparable​​ to the state of​​​‌ the art on multiple​ downstream tasks while delivering​‌ significant speed-ups. Notably, we​​ obtain competitive segmentation results​​​‌ using generic DINOv2 features,​ despite DINOv2 not being​‌ trained on millions of​​ annotated segmentation masks like​​​‌ SAM. When applied to​ CLIP features, our method​‌ demonstrates strong performance in​​ open-vocabulary object detection tasks,​​​‌ highlighting the versatility of​ our approach.

Figure

Cluster​​​‌ and Predict Latent Patches​ for Improved Masked Image​‌ Modeling

Participants: Maxime Oquab​​, Federico Baldassarre,​​​‌ Timothee Darcet, Julien​ Mairal, Piotr Bojanowski​‌.

Masked Image Modeling​​ (MIM) offers a promising​​​‌ approach to self-supervised representation​ learning, however existing MIM​‌ models still lag behind​​ the state-of-the-art. In this​​​‌ paper 8, we​ systematically analyze target representations,​‌ loss functions, and architectures,​​ to introduce CAPI -​​​‌ a novel pure-MIM framework​ that relies on the​‌ prediction of latent clusterings.​​ Our approach leverages a​​​‌ clustering-based loss, which is​ stable to train, and​‌ exhibits promising scaling properties.​​ Our ViT-L backbone, CAPI,​​​‌ achieves 83.8% accuracy on​ ImageNet and 32.1% mIoU​‌ on ADE20K with simple​​ linear probes, substantially outperforming​​​‌ previous MIM methods and​ approaching the performance of​‌ the current state-of-the-art, DINOv2.​​ The approach is illustrated​​​‌ in Figure 5.​

Figure

Entropy Rectifying Guidance​​​‌ for Diffusion and Flow​ Models

Participants: Tariq Berrada​‌ Ifriqi, Adriana Romero-Soriano​​, Michal Drozdzal,​​​‌ Jakob Verbeek, Karteek​ Alahari.

Guidance techniques​‌ are commonly used in​​ diffusion and flow models​​​‌ to improve image quality​ and input consistency for​‌ conditional generative tasks such​​ as class- conditional and​​​‌ text-to-image generation. In particular,​ classifier-free guidance (CFG) is​‌ the most widely adopted​​ guidance technique. It results,​​​‌ however, in trade-offs across​ quality, diversity and consistency:​‌ improving some at the​​ expense of others. While​​​‌ recent work has shown​ that it is possible​‌ to disentangle these factors​​ to some extent, such​​​‌ methods come with an​ overhead of requiring an​‌ additional (weaker) model, or​​ require more forward passes​​​‌ per sampling step. In​ this work 29,​‌ we propose Entropy Rectifying​​ Guidance (ERG), a simple​​​‌ and effective guidance method​ based on inference-time changes​‌ in the attention mechanism​​ of state-of-the-art diffusion transformer​​​‌ architectures, which allows for​ simultaneous improvements over image​‌ quality, diversity and prompt​​ consistency. ERG is more​​​‌ general than CFG and​ similar guidance techniques, as​‌ it extends to unconditional​​ sampling. We show that​​​‌ ERG results in significant​ improvements in various tasks,​‌ including text-to-image, class-conditional and​​ unconditional image generation (see​​​‌ examples in Figure 6​). We also show​‌ that ERG can be​​ seamlessly combined with other​​​‌ recent guidance methods such​ as CADS and APG,​‌ further improving generation results.​​

Figure

Boosting Latent Diffusion with​​ Perceptual Objectives

Participants: Tariq​​​‌ Berrada Ifriqi, Pietro‌ Astolfi, Melissa Hall‌​‌, Marton Havasi,​​ Yohann Benchetrit, Adriana​​​‌ Romero-Soriano, Karteek Alahari‌, Michal Drozdzal,‌​‌ Jakob Verbeek.

Latent​​ diffusion models (LDMs) power​​​‌ state-of-the-art high-resolution generative image‌ models. LDMs learn the‌​‌ data distribution in the​​ latent space of an​​​‌ autoencoder (AE) and produce‌ images by mapping the‌​‌ generated latents into RGB​​ image space using the​​​‌ AE decoder. While this‌ approach allows for efficient‌​‌ model training and sampling,​​ it induces a disconnect​​​‌ between the training of‌ the diffusion model and‌​‌ the decoder, resulting in​​ a loss of detail​​​‌ in the generated images.‌ To remediate this disconnect,‌​‌ we propose to leverage​​ the internal features of​​​‌ the decoder to define‌ a latent perceptual loss‌​‌ (LPL) 23. This​​ loss encourages the models​​​‌ to create sharper and‌ more realistic images. Our‌​‌ loss can be seamlessly​​ integrated with common autoencoders​​​‌ used in latent diffusion‌ models, and can be‌​‌ applied to different generative​​ modeling paradigms such as​​​‌ DDPM with epsilon and‌ velocity prediction, as well‌​‌ as flow matching. Extensive​​ experiments with models trained​​​‌ on three datasets at‌ 256 and 512 resolution‌​‌ show improved quantitative –​​ with boosts between $\mathrm{6‌}\%$ and $20\%‌$ in FID – and‌​‌ qualitative results when using​​ our perceptual loss (see​​​‌ examples in Figure 7‌.

Figure

Lightweight Structure-Aware‌​‌ Attention for Visual Understanding​​

Participants: Heeseung Kwon,​​​‌ Francisco M. Castro,‌ Manuel J. Marin-Jimenez,‌​‌ Nicolas Guil, Karteek​​ Alahari.

Attention operator​​​‌ has been widely used‌ as a basic brick‌​‌ in visual understanding since​​ it provides some flexibility​​​‌ through its adjustable kernels.‌ However, this operator suffers‌​‌ from inherent limitations: (1)​​ the attention kernel is​​​‌ not discriminative enough, resulting‌ in high redundancy, and‌​‌ (2) the complexity in​​ computation and memory is​​​‌ quadratic in the sequence‌ length. In this work‌​‌ 13, we propose​​ a novel attention operator,​​​‌ called Lightweight Structure-aware Attention‌ (LiSA), which has a‌​‌ better representation power with​​ log-linear complexity (see Figure​​​‌ 8). Our operator‌ transforms the attention kernels‌​‌ to be more discriminative​​ by learning structural patterns.​​​‌ These structural patterns are‌ encoded by exploiting a‌​‌ set of relative position​​ embeddings (RPEs) as multiplicative​​​‌ weights, thereby improving the‌ representation power of the‌​‌ attention kernels. Additionally, the​​ RPEs are approximated to​​​‌ obtain log-linear complexity. Our‌ experiments and analyses demonstrate‌​‌ that the proposed operator​​ outperforms self-attention and other​​​‌ existing operators, achieving state-of-the-art‌ results on ImageNet-1K and‌​‌ other downstream tasks such​​ as video action recognition​​​‌ on Kinetics-400, object detection‌ and instance segmentation on‌​‌ COCO, and semantic segmentation​​ on ADE-20K.

Figure

Source-free video domain adaptation​​​‌ by learning from noisy​ labels

Participants: Avijit Dasgupta​‌, C. V. Jawahar​​, Karteek Alahari.​​​‌

Despite the progress seen​ in classification methods, current​‌ approaches for handling videos​​ with distribution shifts in​​​‌ source and target domains​ remain source-dependent as they​‌ require access to the​​ source data during the​​​‌ adaptation stage. In this​ paper 9, we​‌ present a self-training based​​ source-free video domain adaptation​​​‌ approach to address this​ challenge by bridging the​‌ gap between the source​​ and the target domains.​​​‌ We use the source​ pre-trained model to generate​‌ pseudo-labels for the target​​ domain samples, which are​​​‌ inevitably noisy. Thus, we​ treat the problem of​‌ source-free video domain adaptation​​ as learning from noisy​​​‌ labels and argue that​ the samples with correct​‌ pseudo-labels can help us​​ in adaptation. To this​​​‌ end, we leverage the​ cross-entropy loss as an​‌ indicator of the correctness​​ of the pseudo-labels and​​​‌ use the resulting small-loss​ samples from the target​‌ domain for fine-tuning the​​ model. We further enhance​​​‌ the adaptation performance by​ implementing a teacher–student (TS)​‌ framework, in which the​​ teacher, which is updated​​​‌ gradually, produces reliable pseudo-labels.​ Meanwhile, the student undergoes​‌ fine-tuning on the target​​ domain videos using these​​​‌ generated pseudo-labels to improve​ its performance. Extensive experimental​‌ evaluations show that our​​ methods, termed as CleanAdapt,​​​‌ CleanAdapt + TS, achieve​ state-of-the-art results, outperforming the​‌ existing approaches on various​​ open datasets. Our source​​​‌ code is publicly available​.

Figure

Flowception:​​ Temporally Expansive Flow Matching​​​‌ for Video Generation

Participants:​ Tariq Berrada Ifriqi,​‌ John Nguyen, Karteek​​ Alahari, Jakob Verbeek​​​‌, Ricky T. Q.​ Chen.

We present​‌ Flowception 46, a​​ novel non-autoregressive and variable-length​​​‌ video generation framework. Flowception​ learns a probability path​‌ that interleaves discrete frame​​ insertions with continuous frame​​​‌ denoising. Compared to autoregressive​ methods, Flowception alleviates error​‌ accumulation/drift as the frame​​ insertion mechanism during sampling​​​‌ serves as an efficient​ compression mechanism to handle​‌ long-term context (see examples​​ in Figure 10).​​​‌ Compared to full-sequence flows,​ our method reduces FLOPs​‌ for training three-fold, while​​ also being more amenable​​​‌ to local attention variants,​ and allowing to learn​‌ the length of videos​​ jointly with their content.​​​‌ Quantitative experimental results show​ improved FVD and VBench​‌ metrics over autoregressive and​​ full-sequence baselines, which is​​​‌ further validated with qualitative​ results. Finally, by learning​‌ to insert and denoise​​ frames in a sequence,​​​‌ Flowception seamlessly integrates different​ tasks such as image-to-video​‌ generation and video interpolation.​​

Figure

Online In-Context Distillation for​​​‌ Low-Resource Vision Language Models‌

Participants: Zhiqi Kang,‌​‌ Rahaf Aljundi, Vaggelis​​ Dorovatas, Karteek Alahari​​​‌.

As the field‌ continues its push for‌​‌ ever more resources, this​​ work turns the spotlight​​​‌ on a critical question:‌ how can vision-language models‌​‌ (VLMs) be adapted to​​ thrive in low-resource, budget-constrained​​​‌ settings? While large VLMs‌ offer strong performance, they‌​‌ are impractical to deploy​​ in such settings. Small​​​‌ VLMs, on the other‌ hand, are efficient but‌​‌ typically require costly fine-tuning​​ to close the performance​​​‌ gap with larger models‌ in the deployment domain.‌​‌ Inspired by the in-context​​ learning framework, we propose​​​‌ an online In-Context Distillation‌ (ICD) method 48,‌​‌ in which a small​​ VLM collaborates with a​​​‌ stronger teacher model at‌ inference time, distilling its‌​‌ knowledge via sparse demonstrations​​ to efficiently bridge the​​​‌ gap between them (see‌ overview in Figure 11‌​‌). Our method is​​ built on an in-depth​​​‌ analysis that identifies the‌ scale and the choice‌​‌ of models for which​​ vision-language ICL is currently​​​‌ feasible, and demonstrates the‌ advantage of ICL over‌​‌ fine-tuning under constrained compute​​ budgets. We enhance our​​​‌ method with a novel‌ cross-modal demonstration selection strategy,‌​‌ teacher test-time scaling to​​ reduce noise, and student​​​‌ uncertainty conditioning to dynamically‌ populate a demonstration pool‌​‌ and minimize teacher queries.​​ Our ICD method significantly​​​‌ boosts the performance of‌ small models (up to‌​‌ $33\%$) using​​ scarce teacher annotations (as​​​‌ low as $4\%‌$), and competes with‌​‌ the teacher's zero-shot performance.​​

Figure

Counterfactual Learning of​​​‌ Stochastic Policies with Continuous‌ Actions

Participants: Houssam Zenati‌​‌, Pierre Gaillard,​​ Julien Mairal.

Counterfactual​​​‌ reasoning from logged data‌ has become increasingly important‌​‌ for many applications such​​ as web advertising or​​​‌ healthcare. In 20,‌ we address the problem‌​‌ of counterfactual learning of​​ stochastic policies with continuous​​​‌ actions, which raises difficult‌ challenges about (i) data‌​‌ modelization, (ii) optimization, and​​ (iii) evaluation on real​​​‌ data. First, we introduce‌ a modeling strategy based‌​‌ on a joint kernel​​ embedding of contexts and​​​‌ actions, illustrated in Figure‌ 12 which overcomes the‌​‌ shortcomings of previous discretization​​ strategies as shown in​​​‌ 9. Second, we empirically‌ show that the optimization‌​‌ aspect of counterfactual learning​​ is more important than​​​‌ previously thought, and we‌ demonstrate the benefits of‌​‌ proximal point algorithms and​​ differentiable estimators. Finally, we​​​‌ propose an evaluation protocol‌ for offline policies in‌​‌ real-world logged systems, which​​ is challenging since policies​​​‌ cannot be replayed on‌ test data, and we‌​‌ release a new large-scale​​ dataset along with multiple​​​‌ synthetic, yet realistic, evaluation‌ setups.

Figure

MAP​​​‌ Estimation with Denoisers: Convergence‌ Rates and Guarantees

Participants:‌​‌ Scott Pesme, Giacomo​​ Meanti, Michael Arbel​​​‌, Julien Mairal.‌

Denoiser models have become‌​‌ powerful tools for inverse​​ problems, enabling the use​​​‌ of pretrained networks to‌ approximate the score of‌​‌ a smoothed prior distribution.​​​‌ These models are often​ used in heuristic iterative​‌ schemes aimed at solving​​ Maximum a Posteriori (MAP)​​​‌ optimisation problems, where the​ proximal operator of the​‌ negative log-prior plays a​​ central role. In practice,​​​‌ this operator is intractable,​ and practitioners plug in​‌ a pretrained denoiser as​​ a surrogate-despite the lack​​​‌ of general theoretical justification​ for this substitution. In​‌ 36, we show​​ that a simple algorithm,​​​‌ closely related to several​ used in practice, provably​‌ converges to the proximal​​ operator under a log-concavity​​​‌ assumption on the prior​ p. We show that​‌ this algorithm can be​​ interpreted as a gradient​​​‌ descent on smoothed proximal​ objectives. Our analysis thus​‌ provides a theoretical foundation​​ for a class of​​​‌ empirically successful but previously​ heuristic methods. This result​‌ is provided in Figure​​ 13.

Figure

Logarithmic Regret for​‌ Unconstrained Submodular Maximization Stochastic​​ Bandit

Participants: Julien Zhou​​​‌, Pierre Gaillard,​ Thibaud Rahier, Julyan​‌ Arbel.

In 40​​, we address the​​​‌ online unconstrained submodular maximization​ problem (Online USM), in​‌ a setting with stochastic​​ bandit feedback. In this​​​‌ framework, a decision-maker receives​ noisy rewards from a​‌ nonmonotone submodular function, taking​​ values in a known​​​‌ bounded interval. This paper​ proposes Double-Greedy - Explore-then-Commit​‌ (DG-ETC), adapting the Double-Greedy​​ approach from the offline​​​‌ and online full-information settings.​ DG-ETC satisfies a $\mathrm{O‌}\left(dl\mathrm{o}g\right(d\mathrm{T‌}\left)\right)$ problem dependent​ upper bound for the​‌ $1/2$-approximate​​ pseudo-regret, as well as​​​‌ a $O\left(\mathrm{d}{T}^{2/\mathrm{3‌}}log{(dT)}^{\mathrm{1‌}/3}\right)$ problem-free​ one at the same​‌ time, outperforming existing approaches.​​ To that end, we​​​‌ introduce a notion of​ hardness for submodular functions,​‌ characterizing how difficult it​​ is to maximize them​​​‌ with this type of​ strategy.

Hardness

Locally​ Adaptive Online Nonparametric Regression​‌

Participants: Paul Liautaud,​​ Pierre Gaillard, Olivier​​​‌ Wintenberger.

In 32​ and 31, We​‌ study online adversarial regression​​ with convex losses against​​​‌ a rich class of​ continuous yet highly irregular​‌ prediction rules, modeled by​​ Besov spaces $B_{\mathrm{p‌},q}^s$ with​ general parameters $1\le ‌p,q\le \infty$ and smoothness $\mathrm{s‌}>\frac{d}{p}$.​ We introduce an adaptive​‌ wavelet-based algorithm that performs​​ sequential prediction without prior​​​‌ knowledge of $(\mathrm{s},p,\mathrm{q‌})$, and establish​​ minimax-optimal regret bounds against​​​‌ any comparator in ${\mathrm{B}}_{p,q}^{\mathrm{s‌}}$. We further design​​ a locally adaptive extension​​​‌ capable of dynamically tracking​ spatially inhomogeneous smoothness. This​‌ adaptive mechanism adjusts the​​ resolution of the predictions​​​‌ over both time and​ space, yielding refined regret​‌ bounds in terms of​​ local regularity. Consequently, in​​​‌ heterogeneous environments, our adaptive​ guarantees can significantly surpass​‌ those obtained by standard​​ global methods.

Regret

Online Learning​​ Approach for Survival Analysis​​​‌

Participants: Camila Fernandez,‌ Pierre Gaillard, Olivier‌​‌ Wintenberger.

In 10​​, we introduce an​​​‌ online mathematical framework for‌ survival analysis, allowing real‌​‌ time adaptation to dynamic​​ environments and censored data.​​​‌ This framework enables the‌ estimation of event time‌​‌ distributions through an optimal​​ second order online convex​​​‌ optimization algorithm—Online Newton Step‌ (ONS). This approach, previously‌​‌ unexplored, presents substantial advantages,​​ including explicit algorithms with​​​‌ non-asymptotic convergence guarantees. Moreover,‌ we analyze the selection‌​‌ of ONS hyperparameters, which​​ depends on the exp-concavity​​​‌ property and has a‌ significant influence on the‌​‌ regret bound. We introduce​​ an adaptive aggregation method​​​‌ that ensures robustness in‌ hyperparameter selection while maintaining‌​‌ fast regret bounds. These​​ findings can extend beyond​​​‌ the survival analysis field,‌ and are relevant for‌​‌ any case characterized by​​ poor exp-concavity and unstable​​​‌ ONS. Additionally, we propose‌ a stochastic approach for‌​‌ ONS that guarantees logarithmic​​ regret in the case​​​‌ of an exponential hazard‌ model. Next, these assertions‌​‌ are illustrated by simulation​​ experiments, followed by an​​​‌ application to a real‌ dataset. Fernandez et al.‌​‌55 also provides some​​ experimental comparison of existing​​​‌ algorithms for survival analysis.‌

Error

Efficient and Near-Optimal‌ Online Portfolio Selection

Participants:‌​‌ Rémi Jézéquel, Dmitrii​​ Ostrovski, Pierre Gaillard​​​‌.

In 12,‌ we study online portfolio‌​‌ selection as introduced by​​ Cover (1991), where a​​​‌ trader allocates wealth over‌ $d$ assets across $\mathrm{T‌‌}$ rounds to maximize logarithmic​​ return. Cover’s Universal Portfolios​​​‌ achieve worst-case optimal $\mathrm{O‌}(dlog\mathrm{T‌‌})$ regret but require​​ costly $d$-dimensional integration,​​​‌ leading to a prohibitive‌ $\tilde{O}\left({\mathrm{d‌‌}}^4{(T+d)}^{1/‌4}\right)$ per-round runtime.‌ We propose a new‌​‌ algorithm achieving essentially the​​ same regret—up to constants​​​‌ and replacing $log\mathrm{T‌}$ with $log(\mathrm{T‌‌}+d)$—with​​ a drastically improved runtime​​​‌ of $\tilde{O}(‌d^2(\mathrm{T‌‌}+d))$ per round. Our method​​​‌ selects portfolios by minimizing‌ logarithmic loss regularized by‌​‌ a log-determinant barrier, revealing​​ connections between online portfolio​​​‌ selection and classical cutting-plane‌ and interior-point methods.

Online‌​‌ Convex Reinforcement Learning with​​ applications to Demand-Side Management.​​​‌

Participants: Bianca Marin Moreno‌, Khaled Eldowa,‌​‌ Margaux Brégère, Pierre​​ Gaillard, Nadia Oudjane​​​‌.

To counter the‌ challenge of integrating fluctuating‌​‌ renewables into the grid,​​ devices like thermostatically controlled​​​‌ loads (water-heaters, air conditioners,‌ etc) offer flexible demand.‌​‌ However, efficiently controlling a​​ large population of these​​​‌ devices to track desired‌ consumption signals remains a‌​‌ complex challenge. Existing methods​​ lack convergence guarantees and​​​‌ computational efficiency, or resort‌ to regularization techniques instead‌​‌ of tackling the target​​ tracking problem directly. 14​​​‌ addresses these drawbacks. We‌ propose to model the‌​‌ problem as a finite​​ horizon episodic Markov decision​​​‌ process, enabling us to‌ adapt convex optimization algorithms‌​‌ with convergence guarantees and​​​‌ computational efficiency. This framework​ also extends to online​‌ learning scenarios, where daily​​ control decisions are made​​​‌ without prior knowledge of​ consumer behavior and with​‌ daily-changing target profiles due​​ to fluctuations of energy​​​‌ production and inflexible consumption.​ We introduce a new​‌ algorithm, called Online Target​​ Tracker (OTT), the first​​​‌ online learning load control​ method, for which we​‌ prove sub-linear regret. We​​ demonstrate our claims with​​​‌ realistic experiments. This combination​ of optimization and learning​‌ lays the groundwork for​​ more dynamic and efficient​​​‌ load control methods. 33​ studies a generalization of​‌ episodic Reinforcement Learning to​​ convex losses that could​​​‌ be applied for Demand-Side​ Management in an unknown​‌ environment. By introducing a​​ reset-free framework called the​​​‌ periodic framework, 49 weakens​ the episodic assumption to​‌ avoid having to reset​​ the population of the​​​‌ devices to the initial​ distribution at every episode.​‌

DSM

Optimized projection-free algorithms for​ online learning: construction and​‌ worst-case analysis

Participants: Julien​​ Weibel, Pierre Gaillard​​​‌, Wouter Koolen,​ Adrien Taylor.

In​‌ 53, we study​​ projection-free algorithms for online​​​‌ learning with linear optimization​ oracles (Frank–Wolfe methods) to​‌ handle constrained decision sets.​​ We propose an optimized​​​‌ variant of an online​ Frank–Wolfe algorithm with a​‌ simple potential-based analysis, and​​ introduce a semidefinite programming​​​‌ framework to jointly design​ and analyze such algorithms.​‌ Our numerical results suggest​​ that no pure online​​​‌ Frank–Wolfe method in this​ model class can achieve​‌ regret better than $\mathrm{O}\left(T^{3/‌4}\right)$ without additional​ assumptions. We further observe​‌ suboptimal constants in existing​​ methods, anytime guarantees of​​​‌ order $O({\mathrm{t}}^{3/4})‌$, and limited benefits​​ from multiple linear optimization​​​‌ steps per round.

Error​

Optimal and Efficient​‌ Algorithms for Multinomial Logistic​​ Bandits

Participants: Pierre Boudart​​​‌, Pierre Gaillard,​ Alessandro Rudi, Aadirupa​‌ Saha.

In 38​​ and 44, we​​​‌ study active online assortment​ optimization with preference feedback,​‌ a framework for modeling​​ user choice and subsetwise​​​‌ utility maximization with applications​ in advertising, online retail,​‌ recommendation systems, and language​​ model fine-tuning. Existing approaches​​​‌ often rely on unrealistic​ assumptions such as strong​‌ reference items or repeated​​ identical assortments. In 38​​​‌, we design efficient​ regret-minimization algorithms that remove​‌ both of these assumptions.​​ In 44, we​​​‌ improve the asymptotic regret​ by a constant that​‌ may be exponentially large​​ in some cases.

MNL​​​‌ regret

Advancing Prompt-Based​ Methods for Replay-Independent General​‌ Continual Learning

Participants: Zhiqi​​ Kang, Liyuan Wang​​​‌, Xingxing Zhang,​ Karteek Alahari.

General​‌ continual learning (GCL) is​​ a broad concept to​​​‌ describe real-world continual learning​ (CL) problems, which are​‌ often characterized by online​​ data streams without distinct​​​‌ transitions between tasks, i.e.,​ blurry task boundaries. Such​‌ requirements result in poor​​ initial performance, limited generalizability,​​ and severe catastrophic forgetting,​​​‌ heavily impacting the effectiveness‌ of mainstream GCL models‌​‌ trained from scratch (see​​ illustration in Figure 20​​​‌). While the use‌ of a frozen pretrained‌​‌ backbone with appropriate prompt​​ tuning can partially address​​​‌ these challenges, such prompt-based‌ methods remain suboptimal for‌​‌ CL of remaining tunable​​ parameters on the fly.​​​‌ In this regard, we‌ propose an innovative approach‌​‌ named MISA (Mask and​​ Initial Session Adaption) to​​​‌ advance prompt-based methods in‌ GCL 30. It‌​‌ includes a forgetting-aware initial​​ session adaption that employs​​​‌ pretraining data to initialize‌ prompt parameters and improve‌​‌ generalizability, as well as​​ a non-parametric logit mask​​​‌ of the output layers‌ to mitigate catastrophic forgetting.‌​‌ Empirical results demonstrate substantial​​ performance gains of our​​​‌ approach compared to recent‌ competitors, especially without a‌​‌ replay buffer (e.g., up​​ to 18.39, 22.06, and​​​‌ 11.96 points performance lead‌ on CIFAR-100, Tiny-ImageNet, and‌​‌ ImageNet-R, respectively). Moreover, our​​ approach features the plug-in​​​‌ nature for prompt-based methods,‌ independence of replay, ease‌​‌ of implementation, and avoidance​​ of CL-relevant hyperparameters, serving​​​‌ as a strong baseline‌ for GCL research. Our‌​‌ source code is publicly​​ available.

Figure

Unified Breakdown Analysis for​​ Byzantine Robust Gossip

Participants:​​​‌ Renaud Gaucher, Aymeric‌ Dieuleveut, Hadrien Hendrikx‌​‌.

Distributed approaches have​​ many computational benefits, but​​​‌ they are vulnerable to‌ attacks from a subset‌​‌ of devices transmitting incorrect​​ information. This work 28​​​‌ investigates Byzantine-resilient algorithms in‌ a decentralized setting, where‌​‌ devices communicate directly with​​ one another. We investigate​​​‌ the notion of breakdown‌ point, and show an‌​‌ upper bound on the​​ number of adversaries that​​​‌ decentralized algorithms can tolerate.‌ This is done through‌​‌ careful study of a​​ specific graph topology, presented​​​‌ in Figure 21.‌ We introduce CG +‌​‌ , an algorithm at​​ the intersection of ClippedGossip​​​‌ and NNA, two popular‌ approaches for robust decentralized‌​‌ learning. CG + meets​​ our upper bound, and​​​‌ thus obtains optimal robustness‌ guarantees, whereas neither of‌​‌ the existing two does.​​ We provide experimental evidence​​​‌ for this gap by‌ presenting an attack tailored‌​‌ to sparse graphs which​​ breaks NNA but against​​​‌ which CG + is‌ robust.

Figure

Byzantine-Robust Gossip:​​​‌ Insights from a Dual‌ Approach

Participants: Renaud Gaucher‌​‌, Aymeric Dieuleveut,​​ Hadrien Hendrikx.

Distributed​​​‌ learning has many computational‌ benefits but is vulnerable‌​‌ to attacks from a​​ subset of devices transmitting​​​‌ incorrect information. This paper‌ 45 investigates Byzantine-resilient algorithms‌​‌ in a decentralized setting,​​ where devices communicate directly​​​‌ in a peer-to-peer manner‌ within a communication network.‌​‌ We leverage the so-called​​ dual approach for decentralized​​​‌ optimization and propose a‌ Byzantine-robust algorithm. We provide‌​‌ convergence guarantees in the​​​‌ average consensus subcase, discuss​ the potential of the​‌ dual approach beyond this​​ subcase, and re-interpret existing​​​‌ algorithms using the dual​ framework, under the general​‌ update rule presented in​​ Figure 22. Lastly,​​​‌ we experimentally show the​ soundness of our method.​‌

Figure

A​‌ Theoretical Framework for Grokking:​​ Interpolation followed by Riemannian​​​‌ Norm Minimisation

Participants: Etienne​ Boursier, Stott Pesme​‌, Radu-Alexandru Dragomir.​​

In 25, we​​​‌ study the dynamics of​ gradient flow with small​‌ weight decay on general​​ training losses $F$.​​​‌ Under mild regularity assumptions​ and assuming convergence of​‌ the unregularised gradient flow,​​ we show that the​​​‌ trajectory with weight decay​ $\lambda$ exhibits a two-phase​‌ behaviour as $\lambda \to 0$. During the​​​‌ initial fast phase, the​ trajectory follows the unregularised​‌ gradient flow and converges​​ to a manifold of​​​‌ critical points of $\mathrm{F}$ . Then, at time​‌ of order $1/\lambda$, the trajectory​​​‌ enters a slow drift​ phase and follows a​‌ Riemannian gradient flow minimising​​ the ${\ell }_2$-norm​​​‌ of the parameters. This​ purely optimisation-based phenomenon offers​‌ a natural explanation for​​ the grokking effect observed​​​‌ in deep learning, where​ the training loss rapidly​‌ reaches zero while the​​ test loss plateaus for​​​‌ an extended period before​ suddenly improving. We argue​‌ that this generalisation jump​​ can be attributed to​​​‌ the slow norm reduction​ induced by weight decay,​‌ as explained by our​​ analysis. We validate this​​​‌ mechanism empirically on several​ synthetic regression tasks. This​‌ mechanism is illustrated in​​ Figure 23.

Figure​​​‌

Flow Matching for​​ Robust Simulation-Based Inference under​​​‌ Model Misspecification

Participants: Pierre-Louis​ Ruhlmann, Pedro Rodrigues​‌, Michael Arbel,​​ Florence Forbes.

Simulation-based​​​‌ inference (SBI) is transforming​ experimental sciences by enabling​‌ parameter estimation in complex​​ non-linear models from simulated​​​‌ data. A persistent challenge,​ however, is model misspecification:​‌ simulators are only approximations​​ of reality, and mismatches​​​‌ between simulated and real​ data can yield biased​‌ or overconfident posteriors. In​​ 51 We address this​​​‌ issue by introducing Flow​ Matching Corrected Posterior Estimation​‌ (FMCPE), a framework that​​ leverages the flow matching​​​‌ paradigm to refine simulation-trained​ posterior estimators using a​‌ small set of real​​ calibration samples, as illustrated​​​‌ in Figure 24.​ Our approach proceeds in​‌ two stages: first, a​​ posterior approximator is trained​​​‌ on abundant simulated data;​ second, flow matching transports​‌ its predictions toward the​​ true posterior supported by​​​‌ real observations, without requiring​ explicit knowledge of the​‌ misspecification. This design enables​​ FMCPE to combine the​​​‌ scalability of SBI with​ robustness to distributional shift.​‌ Across synthetic benchmarks and​​ real-world datasets, we show​​​‌ that our proposal consistently​ mitigates the effects of​‌ misspecification, delivering improved inference​​ accuracy and uncertainty calibration​​​‌ compared to standard SBI​ baselines, while remaining computationally​‌ efficient.

Figure

Simulation-based inference​ of yeast centromeres

Participants:​‌ Eloïse Touron, Pedro​​ Rodrigues, Julyan Arbel​​, Nelle Varoquaux,​​​‌ Michael Arbel.

The‌ chromatin folding and the‌​‌ spatial arrangement of chromosomes​​ in the cell play​​​‌ a crucial role in‌ DNA replication and genes‌​‌ expression. An improper chromatin​​ folding could lead to​​​‌ malfunctions and, over time,‌ diseases. For eukaryotes, centromeres‌​‌ are essential for proper​​ chromosome segregation and folding.​​​‌ Despite extensive research using‌ de novo sequencing of‌​‌ genomes and annotation analysis,​​ centromere locations in yeasts​​​‌ remain difficult to infer‌ and are still unknown‌​‌ in most species. Recently,​​ genome-wide chromosome conformation capture​​​‌ coupled with next-generation sequencing‌ (Hi-C) has become one‌​‌ of the leading methods​​ to investigate chromosome structures.​​​‌ Some recent studies have‌ used Hi-C data to‌​‌ give a point estimate​​ of each centromere, but​​​‌ those approaches highly rely‌ on a good pre-localization.‌​‌ In 39, we​​ present a novel approach​​​‌ that infers in a‌ stochastic manner the locations‌​‌ of all centromeres in​​ budding yeast based on​​​‌ both the experimental Hi-C‌ map and simulated contact‌​‌ maps using a neural​​ network model as illustrated​​​‌ in Figure 25.‌

Figure

Dual Perspectives on​​​‌ Non-Contrastive Self-Supervised Learning

Participants:‌ Jean Ponce, Basile‌​‌ Terver, Martial Hebert​​, Michael Arbel.​​​‌

The stop gradient and‌ exponential moving average iterative‌​‌ procedures are commonly used​​ in non-contrastive approaches to​​​‌ self-supervised learning to avoid‌ representation collapse, with excellent‌​‌ performance in downstream applications​​ in practice. In 50​​​‌, we investigate these‌ procedures from the dual‌​‌ viewpoints of optimization and​​ dynamical systems. We show​​​‌ that, in general, although‌ they do not optimize‌​‌ the original objective, or​​ any other smooth function,​​​‌ they do avoid collapse.‌ Following prior work, but‌​‌ without any of the​​ extra assumptions used in​​​‌ their proofs, we then‌ show using a dynamical‌​‌ system perspective that, in​​ the linear case, minimizing​​​‌ the original objective function‌ without the use of‌​‌ a stop gradient or​​ exponential moving average always​​​‌ leads to collapse, as‌ shown in Figure 26‌​‌. Conversely, we characterize​​ explicitly the equilibria of​​​‌ the dynamical systems associated‌ with these two procedures‌​‌ in this linear setting​​ as algebraic varieties in​​​‌ their parameter space, and‌ show that they are,‌​‌ in general, asymptotically stable​​. Our theoretical findings​​​‌ are illustrated by empirical‌ experiments with real and‌​‌ synthetic data.

Figure

Learning Theory for Kernel​​​‌ Bilevel Optimization

Participants: Fares‌ El Khoury, Edouard‌​‌ Pauwels, Samuel Vaiter​​, Michael Arbel.​​​‌

Bilevel optimization has emerged‌ as a technique for‌​‌ addressing a wide range​​ of machine learning problems​​​‌ that involve an outer‌ objective implicitly determined by‌​‌ the minimizer of an​​ inner problem. In 27​​​‌, we investigate the‌ generalization properties for kernel‌​‌ bilevel optimization problems where​​ the inner objective is​​​‌ optimized over a Reproducing‌ Kernel Hilbert Space. This‌​‌ setting enables rich function​​ approximation while providing a​​​‌ foundation for rigorous theoretical‌ analysis. In this context,‌​‌ we establish novel generalization​​​‌ error bounds for the​ bilevel problem under finite-sample​‌ approximation. Our approach adopts​​ a functional perspective, inspired​​​‌ by (Petrulionyte et al.,​ 2024), and leverages tools​‌ from empirical process theory​​ and maximal inequalities for​​​‌ degenerate -processes to derive​ uniform error bounds. The​‌ results rely on an​​ equivalence we establish between​​​‌ the estimator implemented in​ practice and an abstract​‌ one derived using the​​ functional perspective that is​​​‌ more amenable to a​ statistical analysis, as shown​‌ in Figure 27.​​ These generalization error estimates​​​‌ allow to characterize the​ statistical accuracy of gradient-based​‌ methods applied to the​​ empirical discretization of the​​​‌ bilevel problem.

Figure

EquiTabPFN:​ A Target-Permutation Equivariant Prior​‌ Fitted Network

Participants: Michael​​ Arbel, David Salinas​​​‌, Frank Hutter.​

Recent foundational models for​‌ tabular data, such as​​ TabPFN, have demonstrated remarkable​​​‌ effectiveness in adapting to​ new tasks through in-context​‌ learning. However, these models​​ overlook a crucial equivariance​​​‌ property: the arbitrary ordering​ of target dimensions should​‌ not influence model predictions.​​ In 22, we​​​‌ identify this oversight as​ a source of incompressible​‌ error, termed the equivariance​​ gap, which introduces instability​​​‌ in predictions. To mitigate​ these issues, we propose​‌ a novel model designed​​ to preserve equivariance across​​​‌ output dimensions, as shown​ in Figure 28.​‌ Our experimental results indicate​​ that our proposed model​​​‌ not only addresses these​ pitfalls effectively but also​‌ achieves competitive benchmark performance.​​

Figure

A New​​ Statistical Model of Star​​​‌ Speckles for Learning to​ Detect and Characterize Exoplanets​‌ in Direct Imaging Observations​​

Participants: Theo Bodrito,​​​‌ Olivier Flasseur, Julien​ Mairal, Jean Ponce​‌, Maud Langlois,​​ Anne-Marie Lagrange.

The​​​‌ search for exoplanets is​ an active field in​‌ astronomy, with direct imaging​​ as one of the​​​‌ most challenging methods due​ to faint exoplanet signals​‌ buried within stronger residual​​ starlight. Successful detection requires​​​‌ advanced image processing to​ separate the exoplanet signal​‌ from this nuisance component.​​ The paper 24 presents​​​‌ a novel statistical model​ that captures nuisance fluctuations​‌ using a multiscale approach,​​ leveraging problem symmetries and​​​‌ a joint spectral channel​ representation grounded in physical​‌ principles. Our model integrates​​ into an interpretable, end-to-end​​​‌ learnable framework for simultaneous​ exoplanet detection and flux​‌ estimation. The proposed algorithm​​ is evaluated against the​​​‌ state of the art​ using datasets from the​‌ SPHERE instrument operating at​​ the Very Large Telescope​​​‌ (VLT). It significantly improves​ the precision-recall tradeoff, notably​‌ on challenging datasets that​​ are otherwise unusable by​​​‌ astronomers. The proposed approach​ is computationally efficient, robust​‌ to varying data quality,​​ and well suited for​​​‌ large-scale observational surveys. The​ model is illustrated in​‌ Figure 29.

Figure​​

Unsupervised Imaging Inverse Problems​‌ with Diffusion Distribution Matching​​

Participants: Giacomo Meanti,​​​‌ Thomas Ryckeboer, Michael​ Arbel, Julien Mairal​‌.

This work 35​​ addresses image restoration tasks​​ through the lens of​​​‌ inverse problems using unpaired‌ datasets. In contrast to‌​‌ traditional approaches—which typically assume​​ full knowledge of the​​​‌ forward model or access‌ to paired degraded and‌​‌ ground-truth images—the proposed method​​ operates under minimal assumptions​​​‌ and relies only on‌ small, unpaired datasets. This‌​‌ makes it particularly well-suited​​ for real-world scenarios, where​​​‌ the forward model is‌ often unknown or mis-specified,‌​‌ and collecting paired data​​ is costly or infeasible.​​​‌ The method leverages conditional‌ flow matching to model‌​‌ the distribution of degraded​​ observations, while simultaneously learning​​​‌ the forward model via‌ a distribution-matching loss that‌​‌ arises naturally from the​​ framework. Empirically, it outperforms​​​‌ both single-image blind and‌ unsupervised approaches on deblurring‌​‌ and non-uniform point spread​​ function (PSF) calibration tasks.​​​‌ It also matches state-of-the-art‌ performance on blind super-resolution.‌​‌ We also showcase the​​ effectiveness of our method​​​‌ with a proof of‌ concept for lens calibration:‌​‌ a real-world application traditionally​​ requiring timeconsuming experiments and​​​‌ specialized equipment. In contrast,‌ our approach achieves this‌​‌ with minimal data acquisition​​ effort. This approach is​​​‌ illustrated in Figure 30‌.

Figure

Optimal transport‌ unlocks end-to-end learning for‌​‌ single-molecule localization

Participants: Romain​​ seailles, Jean-Baptiste Masson​​​‌, Jean Ponce,‌ Julien Mairal.

Single-molecule‌​‌ localization microscopy (SMLM) allows​​ reconstructing biology-relevant structures beyond​​​‌ the diffraction limit by‌ detecting and localizing individual‌​‌ fluorophores – fluorescent molecules​​ stained onto the observed​​​‌ specimen – over time‌ to reconstruct super-resolved images.‌​‌ Currently, efficient SMLM requires​​ non-overlapping emitting fluorophores, leading​​​‌ to long acquisition times‌ that hinders live-cell imaging.‌​‌ Recent deep-learning approaches can​​ handle denser emissions, but​​​‌ they rely on variants‌ of non-maximum suppression (NMS)‌​‌ layers, which are unfortunately​​ non-differentiable and may discard​​​‌ true positives with their‌ local fusion strategy. In‌​‌ this work 52,​​ we reformulate the SMLM​​​‌ training objective as a‌ set-matching problem, deriving an‌​‌ optimal-transport loss that eliminates​​ the need for NMS​​​‌ during inference and enables‌ end-to-end training. Additionally, we‌​‌ propose an iterative neural​​ network that integrates knowledge​​​‌ of the microscope's optical‌ system inside our model.‌​‌ Experiments on synthetic benchmarks​​ and real biological data​​​‌ show that both our‌ new loss function and‌​‌ architecture surpass the state​​ of the art at​​​‌ moderate and high emitter‌ densities. This approach is‌​‌ illustrated in Figure 31​​.

Figure

SpectralEarth: Training‌​‌ Hyperspectral Foundation Models at​​ Scale

Participants: Nassim Ait​​​‌ Ali Braham, C.‌ Albrecht, Julien Mairal‌​‌, Jocelyn Chanussot,​​ Y Wang, Xiao​​​‌ Xiang Zhu.

Foundation‌ models have triggered a‌​‌ paradigm shift in computer​​ vision and are increasingly​​​‌ being adopted in remote‌ sensing, particularly for multispectral‌​‌ imagery. Yet, their potential​​ in hyperspectral imaging (HSI)​​​‌ remains untapped due to‌ the absence of comprehensive‌​‌ and globally representative hyperspectral​​ datasets. To close this​​​‌ gap, in 4 we‌ introduce SpectralEarth, a large-scale‌​‌ multitemporal dataset designed to​​ pretrain hyperspectral foundation models​​​‌ leveraging data from the‌ environmental mapping and analysis‌​‌ program (EnMAP). SpectralEarth comprises​​​‌ 538 974 image patches​ covering 415 153 unique​‌ locations from 11 636​​ globally distributed EnMAP scenes​​​‌ spanning two years of​ archive. In addition, $\mathrm{17‌}.5\%$ of​​ these locations include multiple​​​‌ timestamps, enabling multitemporal HSI​ analysis. Utilizing state-of-the-art selfsupervised​‌ learning algorithms, we pretrain​​ a series of foundation​​​‌ models on SpectralEarth, integrating​ a spectral adapter into​‌ classical vision backbones to​​ accommodate the unique characteristics​​​‌ of HSI. In tandem,​ we construct nine downstream​‌ datasets for land-cover, crop-type​​ mapping, and tree-species classification,​​​‌ providing benchmarks for model​ evaluation. Experimental results support​‌ the versatility of our​​ models and their generalizability​​​‌ across different tasks and​ sensors. We also highlight​‌ computational efficiency during model​​ fine-tuning. In Figure 32​​​‌, we compare the​ size of various datasets​‌ published for Earth observation.​​

Figure

MicroFlow:​‌ Domain-Specific Optical Flow for​​ Ground Deformation Estimation in​​​‌ Seismic Events

Participants: Juliette​ Bertrand, Sophie Giffard-Roisin​‌, James Hollingsworth,​​ Julien Mairal.

Dense​​​‌ ground displacement measurements are​ crucial for geological studies​‌ but are impractical to​​ collect directly. Traditionally, displacement​​​‌ fields are estimated using​ patch matching on optical​‌ satellite images from different​​ acquisition times. While deep​​​‌ learning-based optical flow models​ are promising, their adoption​‌ in ground deformation analysis​​ is hindered by challenges​​​‌ such as the absence​ of real ground truth,​‌ the need for sub-pixel​​ precision, and temporal variations​​​‌ due to geological or​ anthropogenic changes. In particular,​‌ we identify that deep​​ learning models relying on​​​‌ explicit correlation layers struggle​ at estimating small displacements​‌ in real-world conditions. Instead,​​ we propose a model​​​‌ that employs iterative refinements​ with explicit warping layers​‌ and a correlation-independent backbone,​​ enabling sub-pixel precision. Additionally,​​​‌ a non-convex variant of​ Total Variation regularization preserves​‌ fault-line sharpness while maintaining​​ smoothness elsewhere. Our model​​​‌ significantly outperforms widely used​ geophysics methods on semi-synthetic​‌ benchmarks and generalizes well​​ to challenging real-world scenarios​​​‌ captured by both medium-​ and high-resolution sensors. This​‌ work is available in​​ the paper 43 and​​​‌ is illustrated in Figure​ 33.

Figure

Leveraging​​​‌ very high resolution optical​ remote sensing data and​‌ deep learning to assess​​ the potential for photovoltaïc​​​‌ energy production in urban​ areas

Participants: Alessia Boccalatte​‌, Jocelyn Chanussot.​​

Convolutional Neural Networks (CNNs)​​​‌ have shown remarkable success​ in remote sensing tasks.​‌ In urban contexts, recent​​ research has utilized CNNs​​​‌ to generate rooftop segmentation​ masks and determine rooftop​‌ section orientation from aerial​​ images. This cost-effective approach​​​‌ is especially valuable for​ large-scale rooftop solar potential​‌ estimations when detailed three-dimensional​​ data is unavailable. This​​​‌ research, published in 3​, introduces SolarMTNet, a​‌ novel multitask dense-prediction network​​ designed for rooftop solar​​​‌ potential prediction using only​ aerial images. Unlike previous​‌ studies that focus on​​ small manually labeled datasets​​​‌ (approximately 2000 scenes) and​ only segment rooftop orientations​‌ while typically assuming constant​​ slopes, SolarMTNet simultaneously segments​​​‌ both orientations and slopes,​ enhancing the accuracy of​‌ solar potential estimations by​​ 40%. SolarMTNet leverages a​​ large, automatically labeled dataset​​​‌ (up to 280000 scenes)‌ created from open-source Swis‌​‌ geospatial and aerial data,​​ significantly improving generalization. The​​​‌ model is trained on‌ rooftop data from the‌​‌ Zurich and Geneva cantons​​ and cross-validated on the​​​‌ Canton of Vaud, Switzerland.‌ The results show a‌​‌ mean Intersection over Union​​ (mIoU) of 0.67 for​​​‌ orientation segmentation and 0.40‌ for slope segmentation. The‌​‌ estimated irradiance exhibits an​​ absolute mean percentage difference​​​‌ of only 5% compared‌ to real solar cadaster‌​‌ data derived from detailed​​ model-based calculations, primarily du​​​‌ to shading issues. Finally,‌ SolarMTNet has also been‌​‌ tested in different geographica​​ areas outside Switzerland (France​​​‌ and Germany), demonstrating consistent‌ performance across diverse regions‌​‌ and pixel resolutions. The​​ quantification of urban solar​​​‌ potential losses from rooftop‌ superstructures via aerial imagery‌​‌ and Convolutional Neural Networks​​ has also been considered​​​‌ 2.

Hyperspectral Pansharpening‌

Participants: Jocelyn Chanussot.‌​‌

Hyperspectral (HS) pansharpening consists​​ of fusing a high-resolution​​​‌ panchromatic (PAN) band and‌ a low-resolution HS image‌​‌ to obtain a new​​ image with high resolution​​​‌ in both the spatial‌ and spectral domains. These‌​‌ remote sensing products are​​ valuable for a wide​​​‌ range of applications, driving‌ ever-growing research efforts. Nonetheless,‌​‌ results still do not​​ meet application demands. In​​​‌ part, this comes from‌ the technical complexity of‌​‌ the task: compared to​​ multispectral (MS) pansharpening, many​​​‌ more bands are involved,‌ in a spectral range‌​‌ only partially covered by​​ the PAN component and​​​‌ with overwhelming noise. However,‌ another major limiting factor‌​‌ is the absence of​​ a comprehensive framework for​​​‌ the rapid development and‌ accurate evaluation of new‌​‌ methods. This article attempts​​ to address this issue.​​​‌ We started by designing‌ a dataset large and‌​‌ diverse enough to allow​​ reliable training (for data-driven​​​‌ methods) and testing of‌ new methods. Then, we‌​‌ selected a set of​​ state-of-the-art methods, following different​​​‌ approaches characterized by promising‌ performance, and reimplemented them‌​‌ in a single PyTorch​​ framework. Finally, we carried​​​‌ out a critical comparative‌ analysis of all methods,‌​‌ using the most accredited​​ quality indicators. The analysis​​​‌ highlights the main limitations‌ of current solutions in‌​‌ terms of spectral/spatial quality​​ and computational efficiency, and​​​‌ it suggests promising research‌ directions 7.

On‌​‌ a related topic, another​​ work presents a critical​​​‌ survey of deep learning‌ in remote sensing image‌​‌ fusion 16.

Probing​​ Synergistic High-Order Interaction for​​​‌ Multi-Modal Image Fusion

Participants:‌ Jocelyn Chanussot.

Multi-modal‌​‌ image fusion aims to​​ generate a fused image​​​‌ by integrating and distinguishing‌ the cross-modality complementary information‌​‌ from multiple source images.​​ While the cross-attention mechanism​​​‌ with global spatial interactions‌ appears promising, it only‌​‌ captures second-order spatial interactions,​​ neglecting higher-order interactions in​​​‌ both spatial and channel‌ dimensions. This limitation hampers‌​‌ the exploitation of synergies​​ between multi-modalities. To bridge​​​‌ this gap, we introduce‌ in 21 a Synergistic‌​‌ High-order Interaction Paradigm (SHIP),​​ designed to systematically investigate​​​‌ spatial fine-grained and global‌ statistics collaborations between the‌​‌ multi-modal images across two​​ fundamental dimensions: 1) Spatial​​​‌ dimension: we construct spatial‌ fine-grained interactions through element-wise‌​‌ multiplication, mathematically equivalent to​​​‌ global interactions, and then​ foster high-order formats by​‌ iteratively aggregating and evolving​​ complementary information, enhancing both​​​‌ efficiency and flexibility. 2)​ Channel dimension: expanding on​‌ channel interactions with first-order​​ statistics (mean), we devise​​​‌ high-order channel interactions to​ facilitate the discernment of​‌ inter-dependencies between source images​​ based on global statistics.​​​‌ We further introduce an​ enhanced version of the​‌ SHIP model, called SHIP++​​ that enhances the cross-modality​​​‌ information interaction representation by​ the cross-order attention evolving​‌ mechanism, cross-order information integration,​​ and residual information memorizing​​​‌ mechanism. Harnessing high-order interactions​ significantly enhances our model’s​‌ ability to exploit multi-modal​​ synergies, leading in superior​​​‌ performance over state-of-the-art alternatives,​ as shown through comprehensive​‌ experiments across various benchmarks​​ in two significant multi-modal​​​‌ image fusion tasks: pan-sharpening,​ and infrared and visible​‌ image fusion.

Fully-Connected Transformer​​ for Multi-Source Image Fusion​​​‌

Participants: Jocelyn Chanussot.​

Multi-source image fusion combines​‌ the information coming from​​ multiple images into one​​​‌ data, thus improving imaging​ quality. This topic has​‌ aroused great interest in​​ the community. How to​​​‌ integrate information from different​ sources is still a​‌ big challenge, although the​​ existing self-attention based transformer​​​‌ methods can capture spatial​ and channel similarities. In​‌ this paper 19,​​ we first discuss the​​​‌ mathematical concepts behind the​ proposed generalized self-attention mechanism,​‌ where the existing self-attentions​​ are considered basic forms.​​​‌ The proposed mechanism employs​ multilinear algebra to drive​‌ the development of a​​ novel fully-connected self-attention (FCSA)​​​‌ method to fully exploit​ local and non-local domain-specific​‌ correlations among multi-source images.​​ Moreover, we propose a​​​‌ multi-source image representation embedding​ it into the FCSA​‌ framework as a non-local​​ prior within an optimization​​​‌ problem. Some different fusion​ problems are unfolded into​‌ the proposed fully-connected transformer​​ fusion network (FC-Former). More​​​‌ specifically, the concept of​ generalized self-attention can promote​‌ the potential development of​​ self-attention. Hence, the FC-Former​​​‌ can be viewed as​ a network model unifying​‌ different fusion tasks. Compared​​ with state-of-the-art methods, the​​​‌ proposed FC-Former method exhibits​ robust and superior performance,​‌ showing its capability of​​ faithfully preserving information.

GeoFlowNet-SAR:​​​‌ Earthquake Displacement Estimation from​ Synthetic Aperture Radar Images​‌

Participants: Jocelyn Chanussot.​​

Displacement estimation using remote​​​‌ sensing images is an​ effective approach for assessing​‌ surface displacement caused by​​ natural disasters like earthquakes​​​‌ and landslides. By employing​ pixel correlation algorithms, high-precision​‌ displacement maps can be​​ generated from images taken​​​‌ before and after surface​ movement. However, traditional methods​‌ often rely on spatial​​ regularization or frequency masking​​​‌ to reduce high-frequency noise,​ which can smooth spatial​‌ details and result in​​ biased displacement estimates, especially​​​‌ near sharp discontinuities typical​ of earthquake surface ruptures.​‌ Moreover, subpixel displacement estimation​​ using synthetic aperture radar​​​‌ (SAR) images remains a​ challenge compared to optical​‌ images, due to the​​ strong impact of speckle​​​‌ noise. This article 18​ presents GeoFlowNet-SAR, an innovative​‌ subpixel displacement estimation method​​ leveraging SAR images. SAR​​​‌ offers advantages thanks to​ all-weather observation and high​‌ penetration, making it suitable​​ for conditions typically challenging​​​‌ for optical systems in​ the visible light spectrum.​‌ This study uses Sentinel-1​​ SAR single look complex​​ (SLC) images with dual-polarization​​​‌ (VV and VH modes)‌ and interferometric wide (IW)‌​‌ swath mode to balance​​ coverage and resolution. By​​​‌ training on simulated displacement‌ datasets with realistic sharp‌​‌ discontinuities, GeoFlowNet-SAR directly predicts​​ surface displacement fields, providing​​​‌ highly efficient, robust, and‌ precise results while overcoming‌​‌ some limitations of traditional​​ methods.The effectiveness of the​​​‌ proposed methodological contribution is‌ first quantitatively demonstrated using‌​‌ synthetic simulated earthquake datasets,​​ including comparisons with state-of-the-art​​​‌ correlation methods. The method‌ is further validated using‌​‌ two real remote sensing​​ images from the 2019​​​‌ Ridgecrest earthquake and from‌ the 2023 Turkey–Syria earthquake.‌​‌ The observed results from​​ these real datasets confirm​​​‌ the effectiveness of GeoFlowNet-SAR‌ in practical applications.

Kolmogorov–Arnold‌​‌ Network for Hyperspectral Change​​ Detection

Participants: Jocelyn Chanussot​​​‌.

Hyperspectral change detection‌ (HCD) techniques to monitor‌​‌ Earth’s surface processes advanced​​ markedly in recent years.​​​‌ Seasonal variations and associated‌ spectral signatures as well‌​‌ as nonlinear noise patterns​​ emanating from sensors and​​​‌ atmospheric sources pose fundamental‌ challenges in HCD. Advanced‌​‌ deep learning models, such​​ as those that leverage​​​‌ convolutional neural networks (3D-Siamese)‌ or transformers (MLP-Mixer), are‌​‌ increasingly employed to address​​ these challenges. However, they​​​‌ often need substantial training‌ data and computational resources.‌​‌ Here, we show that​​ the Kolmogorov–Arnold network (KAN)​​​‌ can enhance HCD capabilities‌ without the excessive training‌​‌ demand of deep networks.​​ The Kolmogorov–Arnold theorem provides​​​‌ the theoretical foundation for‌ our approach, which is‌​‌ particularly well-suited for hyperspectral​​ data analysis by providing​​​‌ a rigorous basis for‌ handling high-dimensional spectral signatures‌​‌ through dimensional reduction and​​ feature extraction. Our architectural​​​‌ design employs this theoretical‌ framework by incorporating specialized‌​‌ neural network layers that​​ mirror the theorem’s compositional​​​‌ structure, thereby facilitating efficient‌ processing of spectral bands.‌​‌ By replacing the linear​​ weighting scheme with learnable​​​‌ nonlinear functions, the Kolmogorov–Arnold‌ network (KAN) provides a‌​‌ unique capability to capture​​ intricate patterns and irregularities​​​‌ in high-dimensional data. Here,‌ we compare five KAN-based‌​‌ architectures and deep learning​​ models such as the​​​‌ MLP-Mixer, 3D-Siamese, dual-branch Siamese‌ spatial–spectral Transformer attention network‌​‌ (DBS3TAN), and the Swin​​ Transformer for HCD and​​​‌ show that the Chebyshev-KAN‌ model, with an average‌​‌ overall accuracy of 97.35%​​ over four real-world benchmark​​​‌ cases, outperforms other models‌ while having a marked‌​‌ lower complexity than the​​ deep learning models. We​​​‌ also show that the‌ choice of fit nonlinear‌​‌ function and model structure​​ is more important than​​​‌ the number of parameters‌ in KAN-based models 15‌​‌.

ECSPLAIN: Explainability Constrained-claSsifier​​ for Pairing the detection​​​‌ and the Localization of‌ moving Areas from SAR‌​‌ INterferograms

Participants: Jocelyn Chanussot​​.

Detecting slope instabilities​​​‌ on synthetic aperture radar‌ (SAR) interferograms using deep‌​‌ learning approaches presents several​​ challenges. This detection task​​​‌ suffers from the lack‌ of transparency of deep‌​‌ networks, the complexity of​​ the input data (i.e.,​​​‌ complex values, sensitivity to‌ distortions, and presence of‌​‌ counterfactuals), and the complexity​​ of the target phenomena​​​‌ (i.e., the variable velocities‌ and the complex underground‌​‌ processes). In this article​​ 5, we propose​​​‌ a new framework called‌ explainability-constrained classifier for pairing‌​‌ the detection and the​​​‌ localization of moving areas​ on interferograms (ECSPLAIN), to​‌ generate decision, localization, and​​ segmentation maps from a​​​‌ single but explainable classifier​ network. It consists of​‌ training a classifier to​​ detect whether an instability​​​‌ is located in the​ patch or not, and​‌ to explain its decision​​ with a class activation​​​‌ map (CAM) that matches​ the actual location of​‌ the instability. Therefore, by​​ using a single classifier​​​‌ network, the framework can​ pair the detection and​‌ the localization of moving​​ areas. Four CAMs are​​​‌ investigated for the training​ of the ECSPLAIN framework.​‌ Experiments on the ISSLIDE​​ dataset show that our​​​‌ proposal achieves better explainability​ than standard a posteriori​‌ CAMs with more than​​ 0.20 points of improvement​​​‌ in terms of Dice​ and IoU scores. It​‌ also allows competitive performance​​ with segmentation-only networks, with​​​‌ only 0.04 points of​ difference in terms of​‌ Dice and intersection over​​ union (IoU) scores. Thus,​​​‌ the proposed method is​ competitive with the most​‌ efficient methods while being​​ lighter, faster, and delivering​​​‌ a decision based on​ a human-like reasoning process.​‌ Finally, the ECSPLAIN framework​​ is applied to enrich​​​‌ the ISSLIDE dataset, discovering​ more than 470 manually​‌ validated slope instabilities over​​ the Alps.

Challenges in​ Non-Polymeric Crystal Structure Prediction:​‌ Why a Geometric, Permutation-Invariant​​ Loss is Needed

Participants:​​​‌ Emmanuel Jehanno, Romain​ Menegaux, Julien Mairal​‌, Sergei Grudinin.​​

Crystalline structure prediction is​​​‌ an essential prerequisite for​ designing materials with targeted​‌ properties. Yet, it is​​ still an open challenge​​​‌ in materials design and​ drug discovery. Despite recent​‌ advances in computational materials​​ science, accurately predicting three-dimensional​​​‌ non-polymeric crystal structures remains​ elusive. In this work​‌ 47, we focus​​ on the molecular assembly​​​‌ problem, where a set​ S of identical rigid​‌ molecules is packed to​​ form a crystalline structure.​​​‌ Such a simplified formulation​ provides a useful approximation​‌ to the actual problem.​​ However, while recent state-of-the-art​​​‌ methods have increasingly adopted​ sophisticated techniques, the underlying​‌ learning objective remains ill-posed.​​ We propose a better​​​‌ formulation that introduces a​ loss function, illustrated in​‌ Figure 34, capturing​​ key geometric molecular properties​​​‌ while ensuring permutation invariance​ over S. Remarkably, we​‌ demonstrate that within this​​ framework, a simple regression​​​‌ model already outperforms prior​ approaches, including flow matching​‌ techniques, on the COD-Cluster17​​ benchmark, a curated non-polymeric​​​‌ subset of the Crystallography​ Open Database (COD).

Figure​‌

10.1.1 Participation in‌​‌ other International Programs

10.2.1 Horizon Europe

10.2.2 Other european​​​‌ programs/initiatives

J. Chanussot is​ involved in a project​‌ funded by the European​​ Space Agency (ESA): ROSE-L​​​‌ in Harmony: EO Data​ Integration for Global Land​‌ Cover and Vegetation Mapping​​ led by the Canadian​​​‌ company C-Core (2025-2028)

10.3.1 ANR​‌ Project BONSAI

Participants: Michael​​ Arbel.

• Project BONSAI​​​‌ is a multi-disciplinary project​ aiming at integrating knowledge​‌ produced by experts, in​​ the form of simulators,​​​‌ into current machine learning​ frameworks through bilevel optimization​‌ for accurate and efficient​​ inference. We address three​​​‌ challenges. The first one​ is to develop a​‌ deep learning-based approach to​​ simulation-based inference that can​​​‌ adapt to data using​ bilevel optimization. A second​‌ challenge is to depoly​​ the methods to real-world​​​‌ problems which have their​ specificities. A third challenge​‌ is to develop bilevel​​ optimization methods that can​​​‌ handle the non-convexity and​ over-parameterization arising from using​‌ deep learning. The principal​​ investigator is Michael Arbel,​​​‌ and the project involves​ participants from Toulouse School​‌ of Economics, TIMC team​​ at UGA and other​​​‌ INRIA teams (Statify). This​ project started in April​‌ 2024.

10.3.2 MIAI chair:​​ Learning Visual Representations from​​​‌ Interaction for Robot Manipulation​ Tasks

Participants: Pia Bideau​‌, Karteek Alahari.​​

• How to grasp an​​​‌ object has been studied​ in computer vision and​‌ robotics and several approaches​​ to this problem exist​​​‌ - either given a​ 3D shape of an​‌ object contact points are​​ determined that lead to​​​‌ a stable hand object​ configuration or an other​‌ line of work aims​​ at reconstructing stable hand​​​‌ object configurations modelling the​ reconstruction process of hand​‌ pose and object pose​​ jointly. In both cases​​​‌ many solutions are possible,​ although a majority might​‌ not be the natural​​ approach that humans would​​​‌ chose - mainly because​ the intention behind the​‌ grasp is omitted. This​​ project aims at learning​​​‌ visual representations from interaction​ that encode activity information.​‌ Encoding such contextual information​​ appears not only to​​​‌ be relevant to synthesise​ feasible grasps furthermore this​‌ is likely to enhance​​ future generalisation skills facilitating​​​‌ adaptation across the same​ activity but different objects​‌ - grasping a cup​​ to pour something into​​​‌ something shares similar motion​ pattern as grasping a​‌ bottle to pour something​​ into something. Inspired by​​​‌ the effectiveness of human​ grasping, we aim at​‌ finding similarly adaptable representations​​ that are capable of​​ guiding complex manipulation skills.​​​‌ To this end we‌ will fuse ideas relying‌​‌ on classical probabilistic modeling​​ of distributions over possible​​​‌ motion trajectories and latent‌ action representations from a‌​‌ conditional variational autoencoder (CVAE).​​ Both of these directions​​​‌ come with complementary strengths‌ and thus provide promising‌​‌ capabilities of modulating the​​ degree of action abstractions​​​‌ at test time to‌ enable both coarse and‌​‌ fine-grained control for real​​ world robot manipulation tasks.​​​‌ The chair is taking‌ place in collaboration with‌​‌ Karteek Alahari, Xavier Alameda-Pineda,​​ and Pierre-Brice Wieber. We​​​‌ have recruited one PhD‌ student and have an‌​‌ intern starting in February​​ 2025.

10.3.3 MIAI Cluster​​​‌ chair: MOnitoring natural Hazards‌ using AI and Remote‌​‌ sensing (MOHAIR)

Participants: Jocelyn​​ Chanussot.

• J. Chanussot​​​‌ is the co-chair, with‌ Sophie Giffard-Roisin (IRD junior‌​‌ researcher, Laboratoire IsTerre) and​​ Yajing Yang (Associate Professor,​​​‌ LISTIC Univ. Savoie Mont-Blanc).‌ This project started in‌​‌ September 2025. It gathers​​ members from 7 different​​​‌ teams of 6 laboratories‌ in Grenoble, Annecy and‌​‌ Clermont-Ferrand.

  Satellite based remote​​ sensing, using a variety​​​‌ of sensing modalities (optical,‌ radar, hyperspectral, lidar) offers‌​‌ a unique source of​​ information to monitor the​​​‌ environment, with fine spatial‌ resolution, wide coverage and‌​‌ frequent revisit. This enables​​ addressing the challenge of​​​‌ natural hazard monitoring and‌ forecasting, which has a‌​‌ significant societal impact. To​​ fully harness the potential​​​‌ of remote sensing data,‌ advanced algorithms in machine‌​‌ learning, deep learning, or​​ more broadly artificial intelligence,​​​‌ must be developed. Gathering‌ an interdisciplinary team of‌​‌ experts, from data science,​​ environmental and Earth sciences,​​​‌ as well as social‌ sciences, this chair will‌​‌ focus on three important​​ topics: forest monitoring, Earth​​​‌ deformation estimation and volcanic‌ inverse modeling. From a‌​‌ methodological point of view,​​ research will be conducted​​​‌ on the analysis of‌ multimodal time series, multimodal‌​‌ deep and graph learning​​ and foundation models.

10.3.4​​​‌ MIAI chair: Fundamentals of‌ Reinforcement Learning

Participants: Pierre‌​‌ Gaillard.

• P. Gaillard​​ is the co-chair, with​​​‌ Bruno Gaujal (LIG, UGA)‌ of this MIAI chair‌​‌ that focuses on developping​​ advanced methodologies for Reinforcement​​​‌ Learning (RL). The project‌ aims to develop new‌​‌ RL algorithms with strong​​ theoretical foundations and practical​​​‌ effectiveness by exploiting the‌ problem's inherent structure. The‌​‌ focus areas include online​​ control of queueing networks,​​​‌ weakly coupled stochastic dynamic‌ systems (sometimes associated with‌​‌ bandits) and parametric learning​​ for adaptive policies. These​​​‌ three approaches to structured‌ learning will be used‌​‌ for innovative applications in​​ energy, cloud computing, and​​​‌ resource allocation.

10.3.5 Deep‌ Red

Participants: Jocelyn Chanussot‌​‌.

• J. Chanussot is​​ the chair of the​​​‌ Deep Red project from‌ the Foundation Grenoble INP‌​‌ under the patronage of​​ Lynred company (2022-2026). The​​​‌ project aims at popularizing‌ the technology of infrared‌​‌ imaging for new usages.​​

10.3.6 PEPR project Numpex​​​‌

Participants: Hadrien Hendrikx.‌

• The 'Numpex' programme's objectives‌​‌ are to design and​​ develop the software building​​​‌ blocks required to equip‌ future 'exascale machines' and‌​‌ to prepare the major​​ application domains that aim​​​‌ to fully exploit the‌ capabilities of such machines‌​‌ for scientific research and​​​‌ industry alike. This project​ is part of France's​‌ response to the next​​ EuroHPC call for expressions​​​‌ of interest (Projet Exascale​ France) in hosting one​‌ of the two major​​ exascale machines planned in​​​‌ Europe for 2024. In​ this way 'Numpex' will​‌ contribute to the creation​​ of a set of​​​‌ tools, software, applications and​ training which will enable​‌ France to remain one​​ of the leaders in​​​‌ the field of international​ competition through its national​‌ Exascale ecosystem that is​​ in step with European​​​‌ strategy.

10.3.7 PEPR project​ Origins

Participants: Julien Mairal​‌.

• Thoth is involved​​ in the axis “Direct​​​‌ imaging and exoplanet characterization”​ of the PEPR Origins.​‌ This is an on-going​​ collaboration with astronomers from​​​‌ Observatoire de Paris and​ Lyon and with the​‌ Willow team.

11.1.1 Scientific events:​​ organisation

11.1.2 Scientific events:​‌ selection

11.1.3 Journal​​

11.1.4 Invited talks​​​‌

• J. Mairal was an‌ invited speaker at the‌​‌ BASP workshop, Villars sur​​ Ollon. Feb. 2025.
• J.​​​‌ Mairal was an invited‌ speaker at the OSKI‌​‌ workshop, Aussois. March 2025.​​
• J. Mairal was an​​​‌ invited speaker at the‌ GDR-IASIS workshop, Lyon. March‌​‌ 2025.
• J. Mairal was​​ an invited speaker at​​​‌ Academie des Sciences (inter-section‌ meeting). June 2025.
• J.‌​‌ Mairal gave an invited​​ seminar at the ELLIS​​​‌ Stuttgart unit. June 2025.‌
• J. Mairal was an‌​‌ invited speaker at the​​ Non-convex optimization: landscapes, dynamics​​​‌ and learning workshop, EPFL,‌ Aug. 2025.
• J. Mairal‌​‌ was an invited speaker​​ at the GDR-IASIS workshop,​​​‌ Paris. Sept. 2025.
• K.‌ Alahari was a keynote‌​‌ speaker at the Inria-Waterloo​​ workshop at Univ. Waterloo,​​​‌ Canada. May 2025
• K.‌ Alahari was an invited‌​‌ speaker at Journées de​​ statistique de la SFdS,​​​‌ Marseille. June 2025.
• K.‌ Alahari was an invited‌​‌ speaker at the Global​​ AI Frontiers Symposium, Seoul,​​​‌ South Korea. Oct. 2025.‌
• K. Alahari was an‌​‌ invited speaker at the​​ Open Science Days@UGA, Grenoble.​​​‌ Nov. 2025.
• K. Alahari‌ was a keynote speaker‌​‌ at the Sfen workshop​​ on apport de l'IA​​​‌ a la science des‌ materiaux pour l'industrie nucleaire,‌​‌ Paris. Dec. 2025.
• S.​​ Pesme was an invited​​​‌ speaker at Journée scientifique‌ du groupe SMAI-SIGMA, Dec.‌​‌ 2025, Paris.
• S. Pesme​​ gave an invited seminar​​​‌ at Centrale Supelec. Nov.‌ 2025.
• S. Pesme gave‌​‌ a talk at the​​ Oberwolfach Mini-Workshop on Probabilistic​​​‌ Perspectives in Neural Network-Based‌ Machine Learning, Oct. 2025,‌​‌ Oberwolfach, Germany
• S. Pesme​​ gave an invited talk​​​‌ at the Workshop sur‌ les modèles génératifs :‌​‌ diffusion, flow matching et​​ leurs applications, Oct. 2025,​​​‌ Lyon.
• S. Pesme gave‌ an invited seminar at‌​‌ Eindhoven University of Technology,​​ Netherlands, Oct 2025
• S.​​​‌ Pesme gave an invited‌ talk at the Workshop‌​‌ on the Statistical Theory​​ of Neural Networks, May​​​‌ 2025, University of Twente,‌ Netherlands.
• J. Marrie gave‌​‌ an invited seminar at​​ Ecole des Ponts, Marne​​​‌ la Vallée, March 2025.‌
• T. Darcet gave an‌​‌ invited talk at the​​ BLISS summer school, TU​​​‌ Berlin. May 2025.
• T.‌ Bodrito gave a talk‌​‌ at the COBREX seminar.​​ Feb. 2025.
• T. Bodrito​​​‌ gave a talk at‌ the Journées de la‌​‌ Société Française d'Astronomie (SF2A).​​ July 2025.
• P. Gaillard​​​‌ was an invited speaker‌ at a scientific seminar‌​‌ organized by the LabEx​​ EnergyAlps. May 2025.
• H.​​​‌ Hendrikx gave an invited‌ seminar at Inria Montpellier,‌​‌ May 2025.
• H. Hendrikx​​ gave a talk at​​​‌ project Redeem (PEPR IA)‌ annual meeting, October 2025.‌​‌
• M. Arbel was an​​​‌ invited speaker at the​ RKHS Seminars, METU, February​‌ 2025.

11.1.5 Scientific expertise​​

• J. Mairal was a​​​‌ member of the Hemholtz​ panel on scientific imaging.​‌
• J. Mairal was a​​ member of the Prairie​​​‌ panel for junior chairs.​
• J. Mairal was a​‌ panel member for the​​ research council of Norway.​​​‌
• K. Alahari was a​ member of the CRCN/ISFP​‌ 2025 recruitment committee at​​ Grenoble.
• P. Gaillard was​​​‌ a reviewer for the​ JCJC call from ANR.​‌
• H. Hendrikx was a​​ panel member for the​​​‌ TSIA call from ANR,​ subcommittee "IA & Environnements,​‌ écosystèmes, ressources biologiques"

11.1.6​​ Research administration

• J. Mairal​​​‌ is a member of​ the scientific committee (COS)​‌ of Inria Grenoble's research​​ center, and also a​​​‌ member of the scientific​ committee of MIAI.
• K.​‌ Alahari is the deputy​​ scientific director in charge​​​‌ of AI at Inria.​
• K. Alahari is one​‌ of the scientific directors​​ of the PEPR IA​​​‌ national research programme.
• K.​ Alahari is responsible for​‌ the Mathematics and Computer​​ Science specialist field at​​​‌ the MSTII doctoral school.​
• K. Alahari is a​‌ member of commission prospection​​ postes at LJK.
• H.​​​‌ Hendrikx is Chargé de​ mission Science Environnement Société​‌ (SEnS) for Inria Grenoble.​​
• H. Hendrikx is the​​​‌ Inria Transformation Ecologique (TREC)​ representative at UGA.
• H.​‌ Hendrikx is a leading​​ member of the Inria​​​‌ Grenoble socio-environmental roadmap.
• J.​ Chanussot is a member​‌ of the Commission Recherche​​, University Grenoble Alpes.​​​‌

11.2.1​​ Supervision

• Théo Bodrito defended​​​‌ his Phd in June​ 2025. He was co-advised​‌ by Olivier Flasseur, Jean​​ Ponce and Julien Mairal.​​​‌ See the manuscript 41​.
• Timothée Darcet defended​‌ his Phd in June​​ 2025. He was co-advised​​​‌ by Piotr Bojanowski, Maxim​ Oquab, and Julien Mairal.​‌ See the manuscript 42​​
• Juliette Marrie defended her​​​‌ PhD in June 2025.​ She was co-advised by​‌ Michael Arbel, Diane Larlus​​ and Julien Mairal.
• Bianca​​​‌ Marin Moreno defended her​ PhD in October 2025.​‌ She was co-advised by​​ P. Gaillard.
• Zhiqi Kang​​​‌ defended his PhD in​ November 2025. He was​‌ advised by Karteek Alahari.​​
• Anandaramane Candassamy,defended his PhD​​​‌ in september 2025. He​ was co-advised by J.​‌ Chanussot.
• Colin Prieur defended​​ his PhD in november​​​‌ 2025. He was co-advised​ by J. Chanussot

11.2.2​‌ Juries

• J. Mairal was​​ reviewer for the PhD​​​‌ thesis of Samuel Gruffaz,​ Univ. Paris Saclay. 2025.​‌
• J. Mairal was reviewer​​ for the HdR of​​​‌ Thomas Moreau, Univ. Paris​ Saclay. 2025.
• J. Mairal​‌ was a member of​​ the PhD committee of​​​‌ Gaspard Dussert, Univ. Lyon​ 1. 2025.
• J. Mairal​‌ was a member of​​ the HdR commitee of​​​‌ Maxime Sangnier, PSL Sorbonne​ Université. 2025.
• K. Alahari​‌ was a member of​​ the PhD jury of​​​‌ Mohammmed-Yasser Benigmim, IP Paris.​ 2025.
• K. Alahari was​‌ a reviewer for the​​ PhD thesis of Corentin​​​‌ Sautier, Ecole des Ponts​ ParisTech. 2025.
• K. Alahari​‌ was a reviewer for​​ the PhD thesis of​​​‌ Tanay Agrawal, Université Côte​ d'Azur. 2025.
• K. Alahari​‌ was the president of​​ the PhD jury of​​ Guillaume Déau, Univ. Poitiers.​​​‌ 2025.
• P. Gaillard was‌ reviewer for the PhD‌​‌ thesis of Lukas Zierahn,​​ Politecnico di Torino, Italy.​​​‌ 2025.
• P. Gaillard was‌ a member of the‌​‌ PhD committee of Antoine​​ Picard, Univ. Lille. 2025.​​​‌
• J. Chanussot was a‌ reviewer for the PhD‌​‌ of Liang Zhao, University​​ of South Australia (Australia)​​​‌ 2025.
• J. Chanussot was‌ a reviewer for the‌​‌ PhD of Kimmo Riihiaho,​​ University of Jyväskylä (Finland)​​​‌ 2025.
• J. Chanussot was‌ a reviewer for the‌​‌ PhD of Dan Pineau,​​ Université Paris-Saclay, 2025.
• J.​​​‌ Chanussot was a reviewer‌ for the PhD of‌​‌ Yi Wang, TU Munich​​ (Germany) 2025.
• J. Chanussot​​​‌ was a reviewer for‌ the PhD of Sai‌​‌ Reddy B., GITAM -​​ Deemed to be University​​​‌ (India) 2025.
• J. Chanussot‌ was a reviewer for‌​‌ the PhD of Triem​​ Pham, Université Paris-Saclay, 2025.​​​‌
• J. Chanussot was the‌ president of the PhD‌​‌ jury of Astrid Tazzioli,​​ Université PSL Paris, 2025.​​​‌
• J. Chanussot was a‌ reviewer for the PhD‌​‌ of Ritu Yadav, KTH​​ (Sweden), 2025
• J. Chanussot​​​‌ was a reviewer and‌ the president of the‌​‌ committee for the PhD​​ of Vadim Becquet, Université​​​‌ Paris PSL - Mines‌ de Paris, 2025
• J.‌​‌ Chanussot was a reviewer​​ for the HdR of​​​‌ Minh-Tan PHAM, Université de‌ Bretagne Sud, 2025
• M.‌​‌ Arbel was a member​​ of the PhD committee​​​‌ of Alessandro Pasqui, PSL‌ Université de Paris, 2025.‌​‌

11.2.3 Educational and pedagogical​​ outreach

• Master: M. Arbel​​​‌ and J. Mairal, Kernel‌ methods for statistical learning,‌​‌ 36h eqTD, M2, ENS​​ Paris-Saclay/PSL, France.
• Master: M.​​​‌ Arbel, J. Mairal and‌ S. Pesme, From Basic‌​‌ Machine Learning models to​​ Advanced Kernel Learning, 54h​​​‌ eqTD, M2, UGA, Grenoble.‌
• Master: P. Gaillard, Sequential‌​‌ Learning, 12h eqTD, M2,​​ MVA, ENS Paris-Saclay, France.​​​‌
• Master: H. Hendrikx, Numerical‌ Optimization, 40h eqTD, M1,‌​‌ UGA, Grenoble
• Master: J.​​ Chanussot, Hyperspectral imaging, 25h​​​‌ eqTD, M2, Grenoble INP‌

11.3.1 Productions‌​‌ (articles, videos, podcasts, serious​​ games, ...)

• K. Alahari​​​‌ participated in a podcast‌ interview for Interstices 54‌​‌.

11.3.2 Participation in​​ Live events

• K. Alahari​​​‌ co-animated the “Café IA"‌ event at Inria Grenoble.‌​‌
• S. Pesme participated to​​ the “Ateliers scolaires les​​​‌ 9 et 10 octobre‌ au sein du parcours‌​‌ "Éclats de sciences" sur​​ le campus de l'Université​​​‌ Grenoble Alpes à Saint-Martin‌ d'Hères”.
• S. Pesme participated‌​‌ two “Café IA”: at​​ Inria (September 30th 2025),​​​‌ and another with Digital‌ League (December 2nd 2025)Talk‌​‌ at the Math Olympiad​​ Ceremony (June 4, 2025,​​​‌ Université Grenoble-Alpes)
• S. Pesme‌ participated to Classroom sessions‌​‌ for the "Semaine des​​ Maths" (March 10–19, 2025,​​​‌ schools in the Grenoble‌ academy)

11.3.3 Others science‌​‌ outreach relevant activities

• S.​​ Pesme was interviewed for​​​‌ the fête de la‌ science 2025.
• J.‌​‌ Chanussot is a member​​ of the scientific advisory​​​‌ board of the établissement‌ public de coopération culturelle‌​‌ « Territoire de Sciences​​ » with its two​​​‌ components: Cosmocité Museum and‌ Grenoble Casemate.
• J. Chanussot‌​‌ organized a half-day event​​ about thermal imaging for​​​‌ the 10th grade students‌ doing their internship at‌​‌ INRIA.

International journals

• 1​‌ articleP.Pia Bideau​​, D.Duc Pham​​​‌, F.Félicie Dhellemmes​, M.Matthew Hansen​‌ and J.Jens Krause​​. Watching Swarm Dynamics​​​‌ from Above: A Framework​ for Advanced Object Tracking​‌ in Drone Videos.​​International Journal of Computer​​​‌ Vision2026. In​ press. HALDOI
• 2​‌ articleA.Alessia Boccalatte​​ and J.Jocelyn Chanussot​​​‌. Quantifying urban solar​ potential losses from rooftop​‌ superstructures via aerial imagery​​ and Convolutional Neural Networks​​​‌.Renewable Energy249​August 2025, 123088​‌HALDOIback to​​ text
• 3 articleA.​​​‌Alessia Boccalatte, A.​Ankit Jha and J.​‌Jocelyn Chanussot. Leveraging​​ large-scale aerial data for​​​‌ accurate urban rooftop solar​ potential estimation via multitask​‌ learning.Solar Energy​​290April 2025,​​​‌ 113336HALDOIback​ to text
• 4 article​‌N. A.Nassim Ait​​ Ali Braham, C.​​​‌ M.Conrad M Albrecht​, J.Julien Mairal​‌, J.Jocelyn Chanussot​​, W.Wang Yi​​​‌ and X. X.Xiao​ Xiang Zhu. SpectralEarth:​‌ Training Hyperspectral Foundation Models​​ at Scale.IEEE​​​‌ Journal of Selected Topics​ in Applied Earth Observations​‌ and Remote Sensing18​​June 2025, 16780-16797​​​‌HALDOIback to​ text
• 5 articleA.​‌Antoine Bralet, A.​​Abdourrahmane Atto, J.​​​‌Jocelyn Chanussot and E.​Emmanuel Trouvé. ECSPLAIN:​‌ Explainability Constrained-claSsifier for Pairing​​ the detection and the​​​‌ Localization of moving Areas​ from SAR INterferograms.​‌IEEE Transactions on Geoscience​​ and Remote Sensing63​​​‌2025, 5217618:1-18HAL​DOIback to text​‌
• 6 articleA.Antoine​​ Bralet, A. M.​​​‌Abdourrahmane M Atto,​ J.Jocelyn Chanussot and​‌ E.Emmanuel Trouvé.​​ Translation-classification loss for SAR​​​‌ image understanding with deep​ learning.Computer Vision​‌ and Image Understanding257​​2025, 104374HAL​​​‌DOI
• 7 articleM.​Matteo Ciotola, G.​‌Giuseppe Guarino, G.​​Gemine Vivone, G.​​​‌Giovanni Poggi, J.​Jocelyn Chanussot, A.​‌Antonio Plaza and G.​​Giuseppe Scarpa. Hyperspectral​​​‌ Pansharpening: Critical review, tools,​ and future perspectives.​‌IEEE geoscience and remote​​ sensing magazine131​​​‌March 2025, 311-338​HALDOIback to​‌ text
• 8 articleT.​​Timothée Darcet, F.​​​‌Frederico Baldassarre, M.​Maxime Oquab, J.​‌Julien Mairal and P.​​Piotr Bojanowski. Cluster​​​‌ and Predict Latent Patches​ for Improved Masked Image​‌ Modeling.Transactions on​​ Machine Learning Research Journal​​​‌June 2025, 1-26​HALback to text​‌
• 9 articleA.Avijit​​ Dasgupta, C.C.V.​​​‌ Jawahar and K.Karteek​ Alahari. Source-free video​‌ domain adaptation by learning​​ from noisy labels.​​​‌Pattern Recognition161May​ 2025, 111328HAL​‌DOIback to text​​
• 10 articleC.Camila​​​‌ Fernandez, P.Pierre​ Gaillard, J.Joseph​‌ de Vilmarest and O.​​Olivier Wintenberger. Online​​​‌ Convex Optimization for Survival​ Analysis: An Adaptive and​‌ Stochastic Approach.Statistical​​ Papers664May​​​‌ 2025, 86HAL​DOIback to text​‌
• 11 articleE. K.​​Elham Kordi Ghasrodashti,​​ P.Peyman Adibi,​​​‌ H.Hossein Karshenas,‌ H. B.Hamidreza Baradaran‌​‌ Kashani and J.Jocelyn​​ Chanussot. Multimodal Image​​​‌ Classification Based on Convolutional‌ Network and Attention-Based Hidden‌​‌ Markov Random Field.​​IEEE Transactions on Geoscience​​​‌ and Remote Sensing63‌2025, 1-14HAL‌​‌DOI
• 12 articleR.​​Rémi Jézéquel, D.​​​‌ M.Dmitrii M. Ostrovskii‌ and P.Pierre Gaillard‌​‌. Efficient and Near-Optimal​​ Online Portfolio Selection.​​​‌Mathematics of Operations Research‌May 2025HALDOI‌​‌back to text
• 13​​ articleH.Heeseung Kwon​​​‌, F. M.Francisco‌ M. Castro, M.‌​‌ J.Manuel J. Marin-Jimenez​​, N.Nicolas Guil​​​‌ and K.Karteek Alahari‌. Lightweight Structure-Aware Attention‌​‌ for Visual Understanding.​​International Journal of Computer​​​‌ Vision133June 2025‌, 6129–6144HALDOI‌​‌back to text
• 14​​ articleB.Bianca Moreno​​​‌, M.Margaux Brégère‌, P.Pierre Gaillard‌​‌ and N.Nadia Oudjane​​. (Online) Convex Optimization​​​‌ for Demand-Side Management: Application‌ to Thermostatically Controlled Loads‌​‌.Journal of Optimization​​ Theory and Applications205​​​‌3April 2025,‌ 43HALDOIback‌​‌ to text
• 15 article​​S.Seyd Teymoor Seydi​​​‌, M.Mojtaba Sadegh‌ and J.Jocelyn Chanussot‌​‌. Kolmogorov–Arnold Network for​​ Hyperspectral Change Detection.​​​‌IEEE Transactions on Geoscience‌ and Remote Sensing63‌​‌2025, 1-15HAL​​DOIback to text​​​‌
• 16 articleG.Gemine‌ Vivone, L.-J.Liang-Jian‌​‌ Deng, S.Shangqi​​ Deng, D.Danfeng​​​‌ Hong, M.Menghui‌ Jiang, C.Chenyu‌​‌ Li, W.Wei​​ Li, H.Huanfeng​​​‌ Shen, X.Xiao‌ Wu, J.-L.Jin-Liang‌​‌ Xiao, J.Jing​​ Yao, M.Mengmeng​​​‌ Zhang, J.Jocelyn‌ Chanussot, S.Salvador‌​‌ García and A.Antonio​​ Plaza. Deep Learning​​​‌ in Remote Sensing Image‌ Fusion: Methods, protocols, data,‌​‌ and future perspectives.​​IEEE geoscience and remote​​​‌ sensing magazine131‌March 2025, 269-310‌​‌HALDOIback to​​ text
• 17 articleP.​​​‌Peng Wang, Z.‌Zhongchen He, B.‌​‌Bo Huang, M.​​ D.Mauro Dalla Mura​​​‌, H.Henry Leung‌ and J.Jocelyn Chanussot‌​‌. VOGTNet: Variational Optimization-Guided​​ Two-Stage Network for Multispectral​​​‌ and Panchromatic Image Fusion‌.IEEE Transactions on‌​‌ Neural Networks and Learning​​ Systems365May​​​‌ 2025, 9268-9282HAL‌DOI
• 18 articleJ.‌​‌Junjie Wang, J.​​James Hollingsworth, E.​​​‌Erwan Pathier, T.‌Tristan Montagnon, W.‌​‌Wei Li, M.​​Mengmeng Zhang, R.​​​‌Ran Tao, J.‌Jocelyn Chanussot and S.‌​‌Sophie Giffard-Roisin. GeoFlowNet-SAR:​​ Earthquake Displacement Estimation from​​​‌ Synthetic Aperture Radar Images‌.IEEE Transactions on‌​‌ Geoscience and Remote Sensing​​November 2025, 1-12​​​‌HALDOIback to‌ text
• 19 articleX.‌​‌Xiao Wu, Z.-H.​​Zi-Han Cao, T.-Z.​​​‌Ting-Zhu Huang, L.-J.‌Liang-Jian Deng, J.‌​‌Jocelyn Chanussot and G.​​Gemine Vivone. Fully-Connected​​​‌ Transformer for Multi-Source Image‌ Fusion.IEEE Transactions‌​‌ on Pattern Analysis and​​ Machine Intelligence473​​​‌March 2025, 2071-2088‌HALDOIback to‌​‌ text
• 20 articleH.​​​‌Houssam Zenati, A.​Alberto Bietti, M.​‌Matthieu Martin, E.​​Eustache Diemert, P.​​​‌Pierre Gaillard and J.​Julien Mairal. Counterfactual​‌ Learning of Stochastic Policies​​ with Continuous Actions.​​​‌Transactions on Machine Learning​ Research JournalMarch 2025​‌HALback to text​​
• 21 articleM.Man​​​‌ Zhou, N.Naishan​ Zheng, X.Xuanhua​‌ He, D.Danfeng​​ Hong and J.Jocelyn​​​‌ Chanussot. Probing Synergistic​ High-Order Interaction for Multi-Modal​‌ Image Fusion.IEEE​​ Transactions on Pattern Analysis​​​‌ and Machine Intelligence47​2February 2025,​‌ 840-857HALDOIback​​ to text

International peer-reviewed​​​‌ conferences

• 22 inproceedingsM.​Michael Arbel, D.​‌David Salinas and F.​​Frank Hutter. EquiTabPFN:​​​‌ A Target-Permutation EquivariantPrior Fitted​ Network.Proceedings in​‌ Advances in Neural Information​​ Processing Systems 38NeurIPS​​​‌ 202538San diego​ (Californie), United States2025​‌HALback to text​​
• 23 inproceedingsT.Tariq​​​‌ Berrada, P.Pietro​ Astolfi, M.Melissa​‌ Hall, M.Marton​​ Havasi, Y.Yohann​​​‌ Benchetrit, A.Adriana​ Romero-Soriano, K.Karteek​‌ Alahari, M.Michal​​ Drozdzal and J.Jakob​​​‌ Verbeek. Boosting Latent​ Diffusion with Perceptual Objectives​‌.ICLR 2025 -​​ 13rd International Conference on​​​‌ Learning RepresentationsSingapore, Singapore​April 2025, 1-28​‌HALback to text​​
• 24 inproceedingsT.Théo​​​‌ Bodrito, O.Olivier​ Flasseur, J.Julien​‌ Mairal, J.Jean​​ Ponce, M.Maud​​​‌ Langlois and A.-M.Anne-Marie​ Lagrange. A New​‌ Statistical Model of Star​​ Speckles for Learning to​​​‌ Detect and Characterize Exoplanets​ in Direct Imaging Observations​‌.CVPR 2025 -​​ IEEE / CVF Conference​​​‌ on Computer Vision and​ Pattern RecognitionNashville, United​‌ StatesIEEE2025,​​ 1-15HALback to​​​‌ text
• 25 inproceedingsE.​Etienne Boursier, S.​‌Scott Pesme and R.-A.​​Radu-Alexandru Dragomir. A​​​‌ Theoretical Framework for Grokking:​ Interpolation followed by Riemannian​‌ Norm Minimisation.Advances​​ in Neural Information Processing​​​‌ SystemsNeurIPS 2025 -​ Neural Information Processing Systems​‌38San Diego, United​​ StatesDecember 2025HAL​​​‌back to text
• 26​ inproceedingsN.Nan Cai​‌ and P.Pia Bideau​​. Active Event Alignment​​​‌ for Monocular Distance Estimation​.WACV 2024 -​‌ IEEE/CVF Winter Conference on​​ Applications of Computer Vision​​​‌Arizona, Tucson, United States​IEEE2025, 2464-2473​‌HALDOIback to​​ text
• 27 inproceedingsF.​​​‌Fares El Khoury,​ E.Edouard Pauwels,​‌ S.Samuel Vaiter and​​ M.Michael Arbel.​​​‌ Learning Theory for Kernel​ Bilevel Optimization.NeurIPS​‌ 2025 – Advances in​​ Neural Information Processing Systems​​​‌NeurIPS 2025 - 39th​ Annual Conference on Neural​‌ Information Processing SystemsSan​​ Diego, United StatesDecember​​​‌ 2025, pp.1-47HAL​back to text
• 28​‌ inproceedingsR.Renaud Gaucher​​, A.Aymeric Dieuleveut​​​‌ and H.Hadrien Hendrikx​. Unified Breakdown Analysis​‌ for Byzantine Robust Gossip​​.Proceedings of Machine​​​‌ Learning ResearchInternational Conference​ on Machine Learning267​‌Vancouver, CanadaJuly 2025​​, 18868-18896HALback​​​‌ to text
• 29 inproceedings​T. B.Tariq Berrada​‌ Ifriqi, A.Adriana​​ Romero-Soriano, M.Michal​​ Drozdzal, J.Jakob​​​‌ Verbeek and K.Karteek‌ Alahari. Entropy Rectifying‌​‌ Guidance for Diffusion and​​ Flow Models.NeurIPS​​​‌ 2025 - Thirty-ninth Conference‌ on Neural Information Processing‌​‌ SystemsSan Diego (CA),​​ United StatesDecember 2025​​​‌, 1-20HALback‌ to text
• 30 inproceedings‌​‌Z.Zhiqi Kang,​​ L.Liyuan Wang,​​​‌ X.Xingxing Zhang and‌ K.Karteek Alahari.‌​‌ Advancing Prompt-Based Methods for​​ Replay-Independent General Continual Learning​​​‌.ICLR 2025 -‌ International Conference on Learning‌​‌ RepresentationsSingapore, Singapore2025​​, 1-19HALback​​​‌ to text
• 31 inproceedings‌P.Paul Liautaud,‌​‌ P.Pierre Gaillard and​​ O.Olivier Wintenberger.​​​‌ Minimax Adaptive Online Nonparametric‌ Regression over Besov Spaces‌​‌.NeurIPS 2025 -​​ 39th Annual Conference on​​​‌ Neural Information Processing Systems‌San Diego, United States‌​‌2025HALback to​​ text
• 32 inproceedingsP.​​​‌Paul Liautaud, P.‌Pierre Gaillard and O.‌​‌Olivier Wintenberger. Minimax-optimal​​ and Locally-adaptive Online Nonparametric​​​‌ Regression.Proceedings of‌ Machine Learning ResearchALT‌​‌ 2025 - 36th International​​ Conference on Algorithmic Learning​​​‌ Theory272Milan, Italy‌February 2025, 702-735‌​‌HALback to text​​
• 33 inproceedingsB.Bianca​​​‌ Marin Moreno, K.‌Khaled Eldowa, P.‌​‌Pierre Gaillard, M.​​Margaux Brégère and N.​​​‌Nadia Oudjane. Online‌ Episodic Convex Reinforcement Learning‌​‌.ICML 2025 -​​ Proceedings of the 42nd​​​‌ International Conference on Machine‌ Learning267Vancouver, Canada‌​‌2025, 44775--44824HAL​​back to text
• 34​​​‌ inproceedingsJ.Juliette Marrie‌, R.Romain Ménégaux‌​‌, M.Michael Arbel​​, D.Diane Larlus​​​‌ and J.Julien Mairal‌. LUDVIG: Learning-free uplifting‌​‌ of 2d visual features​​ to Gaussian splatting scenes​​​‌.ICCV 2025 -‌ International Conference on Computer‌​‌ VisionHonolulu, Hawai'i, United​​ StatesOctober 2025,​​​‌ 1-24HALback to‌ text
• 35 inproceedingsG.‌​‌Giacomo Meanti, T.​​Thomas Ryckeboer, M.​​​‌Michael Arbel and J.‌Julien Mairal. Unsupervised‌​‌ Imaging Inverse Problems with​​ Diffusion Distribution Matching.​​​‌International Conference on Computer‌ Vision (ICCV)Honolulu, United‌​‌ StatesJune 2025HAL​​back to text
• 36​​​‌ inproceedingsS.Scott Pesme‌, G.Giacomo Meanti‌​‌, M.Michael Arbel​​ and J.Julien Mairal​​​‌. MAP Estimation with‌ Denoisers: Convergence Rates and‌​‌ Guarantees.NeurIPS 2025​​ - 39th Annual Conference​​​‌ on Neural Information Processing‌ SystemsSan Diego, United‌​‌ StatesAugust 2025,​​ 1-30HALback to​​​‌ text
• 37 inproceedingsR.‌Runfeng Qu, O.‌​‌Ole Hall, P.​​ K.Pia K Bideau​​​‌, J.Julie Ouerfelli-Ethier‌, M.Martin Rolfs‌​‌, K.Klaus Obermayer​​ and O.Olaf Hellwich​​​‌. Salience-SGG: Enhancing Unbiased‌ Scene Graph Generation with‌​‌ Iterative Salience Estimation.​​WACV 2026 - IEEE/CVF​​​‌ Winter Conference on Applications‌ of Computer VisionArizona,‌​‌ Tucson, United StatesJanuary​​ 2026HALback to​​​‌ text
• 38 inproceedingsA.‌Aadirupa Saha and P.‌​‌Pierre Gaillard. Finally​​ Rank-Breaking Conquers MNL Bandits:​​​‌ Optimal and Efficient Algorithms‌ for MNL Assortment.‌​‌ICLR 2025 - The​​ Thirteenth International Conference on​​​‌ Learning RepresentationsSingapour, Singapore‌April 2025HALback‌​‌ to textback to​​​‌ text
• 39 inproceedingsE.​Eloïse Touron, P.​‌ L.Pedro Luiz Coelho​​ Rodrigues, J.Julyan​​​‌ Arbel, N.Nelle​ Varoquaux and M.Michael​‌ Arbel. Simulation-based inference​​ of yeast centromeres.​​​‌NeurIPS 2025 - 39th​ Conference on Neural Information​‌ Processing Systems Workshop :​​ The 3rd Workshop on​​​‌ Imageomics: Discovering Biological Knowledge​ from Images Using AI.​‌Copenhagen, DenmarkNovember 2025​​, 1-13HALDOI​​​‌back to text
• 40​ inproceedingsJ.Julien Zhou​‌, P.Pierre Gaillard​​, T.Thibaud Rahier​​​‌ and J.Julyan Arbel​. Logarithmic Regret for​‌ Unconstrained Submodular Maximization Stochastic​​ Bandit.Proceedings of​​​‌ Machine Learning ResearchALT​ 2025 - 36th International​‌ Conference on Algorithmic Learning​​ Theory272Milan, Italy​​​‌PMLR2025, 1-25​HALDOIback to​‌ text

Doctoral dissertations and​​ habilitation theses

• 41 thesis​​​‌T.Théo Bodrito.​ Deep learning for exoplanet​‌ detection in high contrast​​ imaging.Ecole normale​​​‌ Supérieur - PSLJune​ 2025HALback to​‌ text
• 42 thesisT.​​Timothée Darcet. Discovering​​​‌ Complex Structures in Images​ using Large-Scale Self-Supervised Learning​‌.Université Grenoble Alpes​​ [2020-....]July 2025HAL​​​‌back to text

Reports​ & preprints

• 43 misc​‌J.Juliette Bertrand,​​ S.Sophie Giffard-Roisin,​​​‌ J.James Hollingsworth and​ J.Julien Mairal.​‌ MicroFlow: Domain-Specific Optical Flow​​ for Ground Deformation Estimation​​​‌ in Seismic Events.​2025HALback to​‌ text
• 44 miscP.​​Pierre Boudart, P.​​​‌Pierre Gaillard and A.​Alessandro Rudi. Enjoying​‌ Non-linearity in Multinomial Logistic​​ Bandits.July 2025​​​‌HALback to text​back to text
• 45​‌ miscR.Renaud Gaucher​​, A.Aymeric Dieuleveut​​​‌ and H.Hadrien Hendrikx​. Byzantine-Robust Gossip: Insights​‌ from a Dual Approach​​.July 2025HAL​​​‌back to text
• 46​ miscT. B.Tariq​‌ Berrada Ifriqi, J.​​John Nguyen, K.​​​‌Karteek Alahari, J.​Jakob Verbeek and R.​‌ T.Ricky T. Q.​​ Chen. Flowception: Temporally​​​‌ Expansive Flow Matching for​ Video Generation.December​‌ 2025HALback to​​ text
• 47 miscE.​​​‌Emmanuel Jehanno, R.​Romain Menegaux, J.​‌Julien Mairal and S.​​Sergei Grudinin. Challenges​​​‌ in Non-Polymeric Crystal Structure​ Prediction: Why a Geometric,​‌ Permutation-Invariant Loss is Needed​​.2025HALback​​​‌ to text
• 48 misc​Z.Zhiqi Kang,​‌ R.Rahaf Aljundi,​​ V.Vaggelis Dorovatas and​​​‌ K.Karteek Alahari.​ Online In-Context Distillation for​‌ Low-Resource Vision Language Models​​.October 2025HAL​​​‌back to text
• 49​ miscB. M.Bianca​‌ Marin Moreno, M.​​Margaux Brégère, P.​​​‌Pierre Gaillard and N.​Nadia Oudjane. Online​‌ Markov Decision Processes with​​ Terminal Law Constraints.​​​‌January 2026HALback​ to text
• 50 misc​‌J.Jean Ponce,​​ B.Basile Terver,​​​‌ M.Martial Hebert and​ M.Michael Arbel.​‌ Dual Perspectives on Non-Contrastive​​ Self-Supervised Learning.October​​​‌ 2025HALback to​ text
• 51 miscP.-L.​‌Pierre-Louis Ruhlmann, P.​​ L.Pedro Luiz Coelho​​​‌ Rodrigues, M.Michael​ Arbel and F.Florence​‌ Forbes. Flow Matching​​ for Robust Simulation-Based Inference​​ under Model Misspecification.​​​‌December 2025HALback‌ to text
• 52 misc‌​‌R.Romain Seailles,​​ J.-B.Jean-Baptiste Masson,​​​‌ J.Jean Ponce and‌ J.Julien Mairal.‌​‌ Optimal transport unlocks end-to-end​​ learning for single-molecule localization​​​‌.2025HALback‌ to text
• 53 misc‌​‌J.Julien Weibel,​​ P.Pierre Gaillard,​​​‌ W. M.Wouter M.‌ Koolen and A.Adrien‌​‌ Taylor. Optimized projection-free​​ algorithms for online learning:​​​‌ construction and worst-case analysis‌.June 2025HAL‌​‌back to text

Scientific​​ popularization

• 54 articleK.​​​‌Karteek Alahari and L.‌Lorenzo Jacques. Des‌​‌ réseaux de neurones artificiels​​ pour aider les robots​​​‌ à comprendre leur environnement‌ [podcast].IntersticesFebruary‌​‌ 2025HALback to​​ text