Portail HAL-PSL
Not a member yet
293219 research outputs found
Sort by
Certified Per-Instance Unlearning Using Individual Sensitivity Bounds
No full textCertified machine unlearning can be achieved via noise injection leading to differential privacy guarantees, where noise is calibrated to worst-case sensitivity. Such conservative calibration often results in performance degradation, limiting practical applicability. In this work, we investigate an alternative approach based on adaptive per-instance noise calibration tailored to the individual contribution of each data point to the learned solution. This raises the following challenge: how can one establish formal unlearning guarantees when the mechanism depends on the specific point to be removed? To define individual data point sensitivities in noisy gradient dynamics, we consider the use of per-instance differential privacy. For ridge regression trained via Langevin dynamics, we derive high-probability per-instance sensitivity bounds, yielding certified unlearning with substantially less noise injection. We corroborate our theoretical findings through experiments in linear settings and provide further empirical evidence on the relevance of the approach in deep learning settings
Riemannian Stochastic Interpolants for Amorphous Particle Systems
No full textModern generative models hold great promise for accelerating diverse tasks involving the simulation of physical systems, but they must be adapted to the specific constraints of each domain. Significant progress has been made for biomolecules and crystalline materials. Here, we address amorphous materials (glasses), which are disordered particle systems lacking atomic periodicity. Sampling equilibrium configurations of glass-forming materials is a notoriously slow and difficult task. This obstacle could be overcome by developing a generative framework capable of producing equilibrium configurations with well-defined likelihoods. In this work, we address this challenge by leveraging an equivariant Riemannian stochastic interpolation framework which combines Riemannian stochastic interpolant and equivariant flow matching. Our method rigorously incorporates periodic boundary conditions and the symmetries of multi-component particle systems, adapting an equivariant graph neural network to operate directly on the torus. Our numerical experiments on model amorphous systems demonstrate that enforcing geometric and symmetry constraints significantly improves generative performance
Water Isotope Model Intercomparison Project (WisoMIP): Present‐Day Climate
No full textInternational audienceAbstract We present the first results of the Water Isotope Model Intercomparison Project (WisoMIP), with Phase 1 focused on modern simulations (1979–2023) from a suite of isotope‐enabled atmospheric general circulation models nudged to ERA5 reanalyzes. Water sources, mixing, and rainout history influence the isotopic composition of vapor and precipitation, making these simulations powerful tools for tracing the global water cycle. By prescribing identical winds, sea surface temperatures, and sea ice conditions, we isolate differences in water isotope behavior across models, controlling for variability in atmospheric dynamics and mean climate. Our analyses show that the ensemble mean best matches observations, as individual model errors cancel out to yield a more accurate representation of Earth's isotope distributions. We also evaluate trends and responses to major climate modes during the recent warming period, highlighting regional and temporal sensitivities in the isotope signals. These diagnostics extend beyond traditional model evaluation metrics (e.g., temperature, precipitation) to reveal uncertainties in physical processes and guide improvements in model parameterizations. The resulting modern nudged ensemble data set serves as a benchmark for isotope‐enabled model development, satellite product comparison, and understanding of water cycle changes in a warming climate. Given its standardized design and broad participation, WisoMIP provides a valuable “isotope reanalysis” product for applications ranging from paleoclimate reconstruction to model tuning. Our work demonstrates the importance of coordinated isotope model evaluation in advancing the use of water isotopes as a diagnostic tool in climate science
Randomized, multicenter Phase III trial of adjuvant chemotherapy with modified FOLFIRINOX versus capecitabine or gemcitabine in patients with resected ampullary adenocarcinoma
No full textInternational audienceBackground: Ampullary adenocarcinoma (AAC) is a rare and aggressive cancer with a 5-year overall survival (OS) rate ranging from 30% to 67% after resection due to a high recurrence rate. Yet, adjuvant therapy's role is still debated. Recent French FFCD-AC cohort study highlighted that adjuvant therapy, can benefit intermediate and high-risk patients. Chemotherapy regimens, include gemcitabine and 5-fluorouracil (5FU) but practices are highly heterogenous due to the low level of evidence. Previous studies suggest that combination chemotherapy, such as mFOLFIRINOX, could offer improved outcomes.Design: PRODIGE 98 - AMPIRINOX trial (NCT06813976) is a multicenter, open-label, randomized phase 3 trial designed to compare the efficacy of adjuvant mFOLFIRINOX versus single-agent chemotherapy (capecitabine or gemcitabine) in patients with resected AAC. Primary outcome is disease free survival and secondary outcomes include overall survival (OS), safety and quality of life. Patients (ages 18-79) must have undergone macroscopically complete (R0/R1) resection of AAC, with exclusion of patients previously treated with chemotherapy, and pT1N0M0 tumors. Ancillary studies will focus on an in-depth molecular profiling of AAC to identify prognostic and predictive biomarkers. AMPIRINOX is currently recruiting and is expected to provide essential data on how to optimize treatment for AAC patients in the coming years
Efficient Crawling for Scalable Web Data Acquisition
No full textInternational audienceJournalistic fact-checking, as well as social or economic research, require analyzing high-quality statistics datasets (SDs, in short). However, retrieving SD corpora at scale may be hard, inefficient, or impossible, depending on how they are published online. To improve open statistics data accessibility, we present a focused Web crawling algorithm that retrieves as many targets, i.e., resources of certain types, as possible, from a given website, in an efficient and scalable way, by crawling (much) less than the full website. We show that optimally solving this problem is intractable, and propose an approach based on reinforcement learning, namely using sleeping bandits. We propose SB-CLASSIFIER, a crawler that efficiently learns which hyperlinks lead to pages that link to many targets, based on the paths leading to the links in their enclosing webpages. Our experiments on websites with millions of webpages show that our crawler is highly efficient, delivering high fractions of a site's targets while crawling only a small part
“La domination et ses technologies. Appréhension des pushbacks aux frontières de l'Union européenne”, communication pour la section thématique Espaces politiques et frontières extérieures de l’UE à l’épreuve de l’enjeu migatoire, Congrès de l'Association Française de Science Politique, 30 juin - 2 juillet 2026.
No full textInternational audienc
“Migration Law and Politics in Contemporary France. Under the Eyes of the Far-Right”, communication durant le panel Migration politics and authoritarianism in Europe pour la International Conference of Europeanists, Dublin, 16-18 juin 2026.
No full textInternational audienc
Simulated parametric study of coherent differential imaging for exoplanet detection with SPHERE
No full textInternational audienceMotivations: Classical high-contrast imaging (ADI, RDI,…) always requires extensive observation time for speckle calibration. We investigate Coherent Differential Imaging (CDI), a powerful alternative that relies on active, user-controlled modulation rather than passive diversity.Method: We implemented the Pair-Wise Probing algorithm within COMPASS simulations to actively modulate optical aberrations in a VLT/SPHERE-like environment.Results: We performed a comprehensive parametric analysis, demonstrated a contrast gain up to a factor of 11 in simulations and identified distinct corrective regimes.Perspective: This work paves the way for CDI as a highly efficient strategy for next-generation instruments like SPHERE+ and the Roman Space Telescope
The JEDI marker as a universal measure of planetary biodiversity
No full textInternational audienceDespite its critical importance in the formation and maintenance of ecosystems and homeostasis on Earth, biodiversity remains a complex and non-unified concept. Consequently, standards for measuring global biodiversity are lacking, hindering our capacity to document Earth’s biota and track its change. Here, we propose the ‘Joint, cellular life-Encompassing DIversity’ ( JEDI ) marker as a simple, effective and quantitative measure to assess and monitor biodiversity. The JEDI marker is a ribosomal RNA gene fragment that can be amplified from all domains of life using a single pair of PCR primers. We demonstrate the applicability and effectiveness of this approach for assessing biodiversity across ecological and biological scales, from holobionts to diverse ecosystems. In addition, we provide an automated bioinformatic workflow to support the standardised and reproducible analysis of the JEDI marker in future studies. While this approach is not free from trade-offs, we argue that its advantages outweigh its limitations by providing a unique, operational and scalable solution that builds on established infrastructure to integrate the microbial majority into biodiversity assessments and provide fundamental insights into organismal dynamics and associations across domains of life. Thus, the JEDI marker approach addresses the urgent need for a universal and standardised framework to effectively measure and monitor biodiversity at planetary scales in an era of profound global change