Archives ouvertes de Paris-Saclay
Not a member yet
272730 research outputs found
Sort by
Nouvelle ère d’observation en cosmologie avec les relevés DESI et Euclid
No full textNational audienceLa cosmologie a vécu une véritable révolution en 1998 avec la découverte de l’accélération de l’expansion de l’Univers. Impossible dans un Univers composé uniquement de matière et de rayonnements, cette accélération nécessite une nouvelle composante, baptisée énergie noire. De nombreux projets ont vu le jour dans l’objectif de caractériser cette mystérieuse énergie noire au moyen de grands relevés spectroscopiques de galaxies, comme le projet au sol DESI ( Dark Energy Spectroscopic Instrument ) et la mission spatiale Euclid. Par des approches différentes mais complémentaires, ces deux expériences devraient apporter des résultats cruciaux sur l’évolution de l’Univers et la nature de ses constituants
Molecular profiling of chemotherapy-resistant breast cancer reveals DNA methylation remodeling associated with the acquisition of paclitaxel resistance
No full textInternational audienceAims: Chemotherapy resistance remains a major challenge in breast cancer (BC) treatment. This study aimed to investigate the role of DNA methylation in this complex process and evaluate the potential of the DNA methyltransferase inhibitor decitabine (DAC) in restoring chemosensitivity.Methods: Paclitaxel (PAC)-and doxorubicin (DOX)-resistant BC cell lines were derived from luminal A (T-47D), triple-negative (MDA-MB-231), and HER2-positive (JIMT-1) models and characterized by molecular profiling and functional assays. The therapeutic effects of DAC and DOX were assessed in MDA-MB-231 xenografts, and integrative analyses of DNA methylation and gene expression identified pathways associated with resistance. Follow-up analyses were performed in PAC-resistant MAS98.12 patient-derived xenografts (PDX) and in clinical samples from the NeoAva trial (NCT00773695). Results: Resistant cells exhibited a slow-cycling phenotype, reduced tumorigenicity, and widespread genomic alterations. PAC-resistant xenografts showed extensive methylation and transcriptomic reprogramming, partly restored by DAC, which increased Ki-67 expression and enhanced DOX responsiveness. In contrast, PDX tumors displayed less pronounced changes, predominantly hypomethylation, indicating distinct resistance mechanisms. Importantly, xenograft-derived CpG signatures stratified NeoAva patients by treatment response.ConclusionsChemoresistance in BC involves extensive genomic and epigenetic remodeling. Although DAC can modulate methylation and tumor phenotype, rational drug combinations will be required to overcome resistance
IA et justice : la voie française: Orientations et doctrines d’emploi de l’intelligence artificielle dans le système judiciaire en France
No full textAt a time when artificial intelligence, and more particularly generative AI, is becoming embedded at the heart of professional practices, the justice system is facing a transformation of unprecedented scale. Tools designed to assist with legal research, drafting, and the analysis of mass litigation, as well as algorithmic systems and large language models, are opening up new perspectives, while simultaneously posing major risks to the very foundations of the judicial institution.This study provides an in-depth analysis of the policy directions and operational doctrines governing the use of artificial intelligence within the French judicial system. Drawing on a cross-analysis of the main institutional reports and documents published between 2024 and 2025 (by Parliament, the Ministry of Justice, the Court of Cassation, the Conseil d’État, and ethical oversight bodies), it highlights the gradual emergence of a distinct “French approach” to AI in justice.The study focuses in particular on the central issue underlying all of these works: how can the potential of artificial intelligence be harnessed to improve the functioning of the public justice service without undermining the fundamental principles of the rule of law? Behind the promises of efficiency lie decisive questions of digital sovereignty, system reliability, data protection, algorithmic transparency, as well as judicial autonomy and the independence of the judicial authority.Structured in four parts, the study first presents an overview of the development of AI within the legal field, clarifying its key concepts and historical evolution. It then offers a detailed review of French institutional publications, before identifying their convergences around a cautious, regulated, and ethically governed deployment of artificial intelligence. Finally, it examines in depth the specific challenge of judicial independence in light of the growing integration of algorithmic systems into judicial processes.At the intersection of law, technology, and ethics, AI and Justice: The French Approach thus provides a structured reading of current public policy choices and helps to clarify the conditions under which the use of artificial intelligence can remain compatible with a human-centred, independent, and trustworthy system of justice.À l’heure où l’intelligence artificielle, et plus particulièrement l’IA générative, s’installe au cœur des pratiques professionnelles, la justice se trouve confrontée à une transformation d’une ampleur inédite. Outils d’aide à la recherche, à la rédaction ou à l’analyse de contentieux de masse, systèmes algorithmiques et grands modèles de langage ouvrent des perspectives nouvelles, tout en faisant peser des risques majeurs sur les fondements mêmes de l’institution judiciaire.Cette étude propose une analyse approfondie des orientations et doctrines d’emploi de l’intelligence artificielle dans le système judiciaire français. À partir d’un examen croisé des principaux rapports et documents institutionnels publiés entre 2024 et 2025 (Parlement, ministère de la Justice, Cour de cassation, Conseil d’État et instances déontologiques), elle met en lumière l’émergence progressive d’une « voie française » de l’IA en justice.L’étude interroge particulièrement la problématique centrale qui traverse l’ensemble de ces travaux : comment tirer parti des potentialités offertes par l’intelligence artificielle pour améliorer le fonctionnement du service public de la justice, sans porter atteinte aux principes fondamentaux de l’État de droit ? Derrière les promesses d’efficience, se posent en effet des questions décisives de souveraineté numérique, de fiabilité des systèmes, de protection des données, de transparence algorithmique, mais aussi d’autonomie du juge et d’indépendance de l’autorité judiciaire.Structurée en quatre parties, l’étude dresse d’abord le panorama du développement de l’IA dans le champ juridique, en clarifiant ses concepts et son évolution historique. Elle propose ensuite une revue détaillée des publications institutionnelles françaises, avant d’en dégager les convergences autour d’un déploiement mesuré, encadré et éthiquement gouverné. Enfin, elle approfondit l’enjeu spécifique de l’indépendance des magistrats, au regard de l’introduction croissante de systèmes algorithmiques dans les processus judiciaires.À la croisée du droit, de la technologie et de l’éthique, IA et justice : la voie française offre ainsi une lecture structurée des choix publics en cours, et contribue à éclairer les conditions d’un usage de l’intelligence artificielle compatible avec une justice humaine, indépendante et digne de confiance
New Insights on Scalar Promotion with the Polyhedral Model
No full textInternational audienceMemory accesses are a well known bottleneck whose impact might be mitigated by using properly the memory hierarchy until registers. In this paper, we address scalar promotion, a technique to turn temporary arrays into a collection of scalar variables to be allocated to registers. We revisit array scalarization in the light of the recent advances of the polyhedral model. We propose a general algorithm for array scalarization and we show a scalarization of stencil computations thanks to a preliminary preprocessing. Our scalarization algorithm operates on the polyhedral intermediate representation and could be plugged in a polyhedral compiler among other passes. In particular, our scalarization algorithm is parametrized by the program schedule, possibly computed by a previous compilation pass. We present a preliminary experimental validation with promising results
Diachronic Stereo Matching for Multi-Date Satellite Imagery
No full textInternational audienceRecent advances in image-based satellite 3D reconstruction have progressed along two complementary directions. On one hand, multi-date approaches using NeRF or Gaussian-splatting jointly model appearance and geometry across many acquisitions, achieving accurate reconstructions on opportunistic imagery with numerous observations. On the other hand, classical stereoscopic reconstruction pipelines deliver robust and scalable results for simultaneous or quasi-simultaneous image pairs. However, when the two images are captured months apart, strong seasonal, illumination, and shadow changes violate standard stereoscopic assumptions, causing existing pipelines to fail. This work presents the first Diachronic Stereo Matching method for satellite imagery, enabling reliable 3D reconstruction from temporally distant pairs. Two advances make this possible: (1) fine-tuning a state-of-the-art deep stereo network that leverages monocular depth priors, and (2) exposing it to a dataset specifically curated to include a diverse set of diachronic image pairs. In particular, we start from a pretrained MonSter model, trained initially on a mix of synthetic and real datasets such as SceneFlow and KITTI, and fine-tune it on a set of stereo pairs derived from the DFC2019 remote sensing challenge. This dataset contains both synchronic and diachronic pairs under diverse seasonal and illumination conditions. Experiments on multi-date WorldView-3 imagery demonstrate that our approach consistently surpasses classical pipelines and unadapted deep stereo models on both synchronic and diachronic settings. Fine-tuning on temporally diverse images, together with monocular priors, proves essential for enabling 3D reconstruction from previously incompatible acquisition dates. Left image (winter) Right image (autumn) DSM geometry Ours (1.23 m) Zero-shot (3.99 m) LiDAR GT Figure 1. Output geometry for a winter-autumn image pair from Omaha (OMA 331 test scene). Our method recovers accurate geometry despite the diachronic nature of the pair, exhibiting strong appearance changes, which cause existing zero-shot methods to fail. Missing values due to perspective shown in black.Mean altitude error in parentheses; lower is better
A forward-only scheme for online learning of proposal distributions in particle filters
No full textWe introduce a new online approach for constructing proposal distributions in particle filters using a forward scheme. Our method progressively incorporates future observations to refine proposals. This is in contrast to backward-scheme algorithms that require access to the entire dataset, such as the iterated auxiliary particle filters (Guarniero et al., 2017, arXiv:1511.06286) and controlled sequential Monte Carlo (Heng et al., 2020, arXiv:1708.08396) which leverage all future observations through backward recursion. In comparison, our forward scheme achieves a gradual improvement of proposals that converges toward the proposal targeted by these backward methods. We show that backward approaches can be numerically unstable even in simple settings. Our forward method, however, offers significantly greater robustness with only a minor trade-off in performance, measured by the variance of the marginal likelihood estimator. Numerical experiments on both simulated and real data illustrate the enhanced stability of our forward approach
SCAR : a self-consistent recurrent cell for real-time finite strain elastoplastic simulations
No full textComplex fabrication and forming processes operating under finite strains could benefit significantly from optimization loops of process parameters, which are often hindered by the prohibitive computational costs of process modeling. Neural networks present a promising solution to derive fast and accurate surrogate models, thereby enabling such optimizations. Furthermore, many processes involve substantial inherent variability, and hence often require manual process control. Neural networks could also provide real-time predictions that would greatly assist in decision-making. Although recursive neural networks have been applied in mechanics, their use in modeling elastoplastic behavior at finite strains remains underexplored. This paper introduces a new family of self-consistent recurrent cells, referred to as SCAR. These cells are specifically designed to address history-dependent problems, such as elastoplasticity, and ensure compliance with key properties required for such applications. To evaluate the SCAR cells, a generic architecture named PlastiNN, featuring a spatially resolved neural decoder, is employed. This approach results in faster training times and more accurate predictions in comparison to commonly used architectures. Additionally, PlastiNN can accommodate a series of successive loads on a workpiece, which is critical for most fabrication and forming processes. The effectiveness of this strategy is demonstrated by comparing SCAR cells to other recurrent cells within the PlastiNN architecture through a comprehensive benchmark including two datasets of 1D and 3D simulations, ranging from challenging toy applications to more realistic industrial test cases. Results highlight the superiority of the proposed recurrent neural network architecture for modeling elastic-plastic behavior at finite strains in engineering processes