Archive ouverte de Centrale Lyon
Not a member yet
32420 research outputs found
Sort by
Dropout-Robust Mechanisms for Differentially Private and Fully Decentralized Mean Estimation
No full textAchieving differentially private computations in decentralized settings poses significant challenges, particularly regarding accuracy, communication cost, and robustness against information leakage. While cryptographic solutions offer promise, they often suffer from high communication overhead or require centralization in the presence of network failures. Conversely, existing fully decentralized approaches typically rely on relaxed adversarial models or pairwise noise cancellation, the latter suffering from substantial accuracy degradation if parties unexpectedly disconnect. In this work, we propose IncA, a new protocol for fully decentralized mean estimation, a widely used primitive in data-intensive processing. Our protocol, which enforces differential privacy, requires no central orchestration and employs low-variance correlated noise, achieved by incrementally injecting sensitive information into the computation. First, we theoretically demonstrate that, when no parties permanently disconnect, our protocol achieves accuracy comparable to that of a centralized setting-already an improvement over most existing decentralized differentially private techniques. Second, we empirically show that our use of low-variance correlated noise significantly mitigates the accuracy loss experienced by existing techniques in the presence of dropouts
Oraux de mathématiques, Concours ENS PC 2025
No full textLicenceCe fichier regroupe un certain nombre d’exercices de mathématiques qui ont été donnés à l’épreuve oral des ENS, concours PC en 2025. Des éléments de réponses sont proposés mais qui ne constituent en rien une correction officielle.</div
Assessment of Three Deep Reinforcement Learning Algorithms for District Heating Network Optimal Control
No full textInternational audienceDistrict Heating Networks (DHNs) play a critical role in reducing carbon emissions and enhancing the sustainability of urban heating systems. Optimizing their performance is essential for achieving energy efficiency and cost-effectiveness. This study investigates the potential of Reinforcement Learning (RL) as a tool for improving the operational control of DHNs, focusing on the dynamic optimization of key parameters under real-world constraints. A dynamic simulation model of the Châteaubriant DHN in France was developed using Simulink, capturing essential system constraints and dynamics. The model was validated against historical data, demonstrating strong agreement between predicted and measured temperature profiles, and subsequently served as a digital twin within an RL-based control framework. Three widely adopted RL algorithms, namely: Twin Delayed Deep Deterministic Policy Gradient (TD3), Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC), were evaluated for their ability to enhance system performance compared to a baseline control strategy. Results demonstrate that RL can significantly improve DHN operations. Among the tested algorithms, TD3 achieved the highest yearly cost savings of 2.85% by reducing fuel consumption and optimizing temperature regulation. PPO and SAC also delivered equivalent savings of 2.33% and 2.11%, respectively. However, SAC required more adjustments to its exploration strategy to achieve such performance. These findings provide valuable insights into each algorithm's control strategies and enable energy system designers and policymakers to choose the proper strategy based on physical constraints
Apprentissage continu adaptatif et efficace dans des environnements dynamiques
No full textAs data in real-world applications continuously evolve, the ability of artificial intelligence systems to learn incrementally while preserving previously acquired knowledge has become increasingly critical. However, deploying continual learning (CL) methods in practice is impeded by blurred task boundaries, severe data imbalance, and the high computational demands and data privacy concerns associated with large models. This thesis addresses these challenges through three core contributions, thus enhancing the feasibility and robustness of CL in dynamic environments. First, to manage blurred task boundaries, where data distributions often overlap, we propose a novel Distribution-Shift Incremental Learning (DS-IL) scenario. In this framework, an entropy-guided learning approach effectively leverages these overlaps to mitigate catastrophic forgetting without maintaining large memory buffers. In real-world scenarios, data imbalance is a common challenge that can significantly hinder the performance of learning systems. To address this issue, our second contribution introduces a Memory Selection and Contrastive Learning (MSCL) strategy. By actively sampling representative instances and coupling them with current data in a contrastive loss, the model better balances underrepresented classes and domains. This approach not only preserves crucial historical information but also maintains robust performance under significantly skewed data distributions. Finally, to alleviate the computational overhead of continually training diffusion models, particularly relevant in scenarios with data privacy constraints or prohibitive storage costs, we introduce a Multi-Mode Adaptive Generative Distillation (MAGD) framework. Using generative distillation, noisy intermediate representations, and exponential moving averages, this method enables efficient continual updates while preserving high-quality image generation and classification performance. Collectively, these contributions form a comprehensive framework for scalable, memory-efficient, and computationally tractable continual learning. Through effective knowledge retention, dynamic adaptation to imbalanced data, and resource-efficient generative replay, this thesis expands the applicability of CL methods to a wider range of real-world settings.Face à l'évolution continue des données dans les applications du monde réel, la capacité des systèmes d'intelligence artificielle à apprendre de manière incrémentale tout en préservant les connaissances acquises précédemment est devenue de plus en plus critique. Cependant, le déploiement des méthodes d'apprentissage continu (CL) dans la pratique est entravé par des frontières de tâches floues, un déséquilibre sévère des données, et les fortes exigences computationnelles et les préoccupations liées à la confidentialité des données associées aux grands modèles. Cette thèse aborde ces défis à travers trois contributions principales, améliorant ainsi la faisabilité et la robustesse du CL dans des environnements dynamiques. Premièrement, pour gérer les frontières de tâches floues, où les distributions de données se chevauchent souvent, nous proposons un nouveau scénario d'Apprentissage Incrémental du Changement de Distribution (DS-IL). Dans ce cadre, une approche d'apprentissage guidée par l'entropie exploite efficacement ces chevauchements pour atténuer l'oubli catastrophique sans maintenir de grands tampons de mémoire. Dans les scénarios réels, le déséquilibre des données est un défi commun qui peut entraver significativement la performance des systèmes d'apprentissage. Pour aborder ce problème, notre deuxième contribution introduit une stratégie de Sélection de Mémoire et d'Apprentissage Contrastif (MSCL). En échantillonnant activement des instances représentatives et en les couplant avec des données actuelles dans une perte contrastive, le modèle équilibre mieux les classes et les domaines sous-représentés. Cette approche préserve non seulement des informations historiques cruciales, mais maintient également une performance robuste sous des distributions de données considérablement biaisées. Enfin, pour alléger la charge computationnelle de la formation continue des modèles de diffusion, particulièrement pertinente dans des scénarios avec des contraintes de confidentialité des données ou des coûts de stockage prohibitifs, nous introduisons un cadre de Distillation Générative Adaptative Multi-Mode (MAGD). Utilisant la distillation générative, des représentations intermédiaires bruyantes, et des moyennes mobiles exponentielles, cette méthode permet des mises à jour continues efficaces tout en préservant une haute qualité de génération d'images et de performance de classification. Collectivement, ces contributions forment un cadre complet pour un apprentissage continu, évolutif, efficace en mémoire et gérable computationnellement. À travers une rétention de connaissances efficace, une adaptation dynamique à des données déséquilibrées et une relecture générative efficiente en ressources, cette thèse étend l'applicabilité des méthodes de CL à un plus large éventail de paramètres du monde réel
Tunable Morphology of GaAs Nanowires by Wet Chemical Etching: Quantum Confinement and 1D Photonic structure
No full textInternational audienc
Flow dynamics of volumes of large light particles released on submerged steep slopes
No full textInternational audienceThe aim of the study is to carry out two-phase gravity driven flows of large light particles on submerged steep slopes and, from their characteristics, identify the criteria that govern the transition between a purely dense and a mixed (dense-suspended) regime. For that, volumes of large light particles are released without any initial velocity at the top of a 2D flume immersed in a 20 m3 tank filled with tap water. The particles are spherical, monodisperse, with two different diameters: 10.6 and 14.4 mm. The volume expansion β, the height H and the front velocity Uf of the heavy flows are investigated varying the volume released between 0.2 L and 3 L, and the tilt angle of the flume (θ) between 30° and 60°. Two different regimes are observed, one where all the particles move in close contact with each other (dense regime), and as the volume and/or flume angle increases, another where part of the particles are suspended (mixed regime). We find that the overall dynamics of the flow is governed by a buoyancy/drag equilibrium ruled by the densimetric Froude number Fr = Uf/(√(∆ρg/ρw)H, with g the gravity acceleration and ∆ρ = ρ - ρw. ρ is the density of the flow and ρw that of the ambient fluid: water. The corresponding drag coefficient exerted on the particles volume is found to vary as Cd ≈ 2Fr-1.6 . The key parameter for the onset of the flow of some of the particles in suspension and the transition to the mixed regime proves to be S tθ = S t cos θ/(1 -S t sin θ), with S t = vs /Uf the Stokes number comparing the settling velocity of the particles to the flow front velocity. While pressure remains hydrostatic within the flow in the dense regime (S tθ > 0.9), it increases as suspension occurs in the mixed case and S tθ decreases. The dynamic pressure at the forehead of the volume then evolves as Ps ≈ ∆ρgHcos(θ)(1.9 - S tθ)2
Evaluation of cancer cells mechanical phenotype associated with resistance to treatment in myeloid leukemia
No full textInternational audienceAcute myeloid leukemia (AML) is a cancer of the myeloid line of blood cells, characterized by the abnormal proliferation of leukemia cells (or blasts) that build up in the bone marrow and the blood. Despite the recent progress in therapies, which consist essentially in intensive cycles of chemotherapy or the use of targeted therapies, most of the AML patients do not recover, having a five-year survival rate of 20%. This poor prognosis may be explained by tumor cell heterogeneity, which could be related to cellular differentiation, as well as to the tumor microenvironment. Indeed, the rapid clonal expansion of leukemic blasts within the bone marrow alters the physical characteristics of the tumor microenvironment and decreases the available space for each cell type. This project seeks to establish a correlation between drug resistance and cell intrinsic stiffness, in the context of the dynamic dialogue between leukemia cells and their microenvironment.One of the most widely utilized passive microfluidic methods in literature for measuring cellular mechanical properties with high throughput involves the monitoring of cell deformations as they flow passively through constricted channels. Here we propose an original readout traducing the way the cell perturbs the pressure distribution within the device, as it blocks the flow when passing through the constriction. Preliminary results suggest different mechanical profile associated with AML cell lines that are sensitive or resistant to chemotherapy, hence demonstrating the pertinence of our approach
Un cadre de substitution unifié pour des informations sur la fiabilité des systèmes mécaniques basées sur les données
No full textModern engineering structures, from simple single degree of freedom (SDOF) systems to complex continuous or multi-degree of freedom (MDOF) configurations and finally to continuous systems, face high-dimensional uncertainties due to variable loads, material properties, and environmental conditions. Traditional approaches, such as Monte Carlo simulation, can become prohibitively expensive when simulating a large parameter space or modeling intricate structural behavior. In this work, we solve these problems by using probabilistic methods with machine learning models. These models include Random Forests, Gradient Boosting, XGBoost, and Neural Networks. They can predict structural responses with less computation. Therefore, it easier to study uncertainty and sensitivity.This paper describes a framework for studying structural dynamics and reliability. It combines theoretical modeling of continuous systems with experimental tests on steel frames. A beam-like structure with infinite degrees of freedom is analyzed using the Rayleigh-Ritz method. Partial differential equations are used to account for mass, stiffness, and damping. Tuned Mass Dampers (TMDs) are added to reduce resonant vibrations. Sampling methods like Latin Hypercube Sampling and Monte Carlo simulation are used to assess reliability under uncertain loads. Machine learning models such as Random Forest and Neural Networks are trained as surrogates to predict failure probabilities. These models show how system parameters interact in nonlinear ways. The framework is tested on a two-storey steel frame in a laboratory. A mechanical shaker applies dynamic forces with different characteristics, such as kurtosis, root mean square (RMS), skewness, and crest factor. Force and acceleration measurements record the structure's responses. Machine learning models, including Random Forest, Gradient Boosting, XGBoost, and Neural Networks, are trained on these features. The test results demonstrated that: 1) Tuned Mass Dampers enhance structural performance and reduce failure risk under dynamic loads. 2) Machine learning models with statistical feature extraction outperform traditional methods and provide practical advantages.Les structures d'ingénierie modernes, des systèmes simples à un seul degré de liberté (SDOF) aux configurations complexes continues ou à plusieurs degrés de liberté (MDOF) et enfin aux systèmes continus, sont confrontées à des incertitudes dimensionnelles élevées en raison des charges variables, des propriétés des matériaux et des conditions environnementales. Les approches traditionnelles, telles que la simulation de Monte Carlo, peuvent devenir extrêmement coûteuses lors de la simulation d'un grand espace de paramètres ou de la modélisation d'un comportement structurel complexe. Dans ce travail, nous résolvons ces problèmes en utilisant des méthodes probabilistes avec des modèles d'apprentissage automatique. Ces modèles incluent Random Forests, Gradient Boosting, XGBoost et Neural Networks. Ils peuvent prédire les réponses structurelles avec moins de calcul. Par conséquent, il est plus facile d'étudier l'incertitude et la sensibilité.Cet article décrit un cadre d'étude de la dynamique et de la fiabilité des structures. Il combine la modélisation théorique des systèmes continus avec des tests expérimentaux sur des cadres en acier. Une structure de type poutre avec des degrés de liberté infinis est analysée à l'aide de la méthode Rayleigh-Ritz. Des équations aux dérivées partielles sont utilisées pour tenir compte de la masse, de la rigidité et de l'amortissement. Des amortisseurs de masse accordés (TMD) sont ajoutés pour réduire les vibrations résonnantes. Des méthodes d'échantillonnage telles que l'échantillonnage par hypercube latin et la simulation de Monte Carlo sont utilisées pour évaluer la fiabilité sous des charges incertaines. Des modèles d'apprentissage automatique tels que Random Forest et Neural Networks sont formés comme substituts pour prédire les probabilités de défaillance. Ces modèles montrent comment les paramètres du système interagissent de manière non linéaire. La structure est testée sur une charpente en acier à deux étages dans un laboratoire. Un vibrateur mécanique applique des forces dynamiques avec différentes caractéristiques, telles que l'aplatissement, la moyenne quadratique (RMS), l'asymétrie et le facteur de crête. Les mesures de force et d'accélération enregistrent les réponses de la structure. Des modèles d'apprentissage automatique, notamment Random Forest, Gradient Boosting, XGBoost et Neural Networks, sont formés sur ces caractéristiques. Les résultats des tests ont démontré que : 1) Les amortisseurs de masse accordés améliorent les performances structurelles et réduisent le risque de défaillance sous des charges dynamiques. 2) Les modèles d'apprentissage automatique avec extraction de caractéristiques statistiques surpassent les méthodes traditionnelles et offrent des avantages pratiques
Projet ciblé VERTIGO (VERTIcal Gan for high vOltage)
No full textLes journées scientifiques annuelles du PEPR électronique sont adossées à C'Nano 2025International audienc
Event knowledge and object-scene knowledge jointly influence fixations in scenes
No full textInternational audienceViewers of real-world scenes typically have knowledge about preceding events ("event knowledge") and the relationships between objects ("object-scene knowledge"). We examined how these knowledge types interact to influence gaze. We recorded eye movements of 48 participants viewing sequences of film frames showing unfolding events. To manipulate event knowledge, we concluded each sequence with an identical "critical frame" that either naturally followed preceding frames or was unrelated. To represent object-scene knowledge, we calculated semantic similarity scores for objects within critical frames using a distributional semantic model. Lower scores indicated objects that were less consistent within a scene. Results showed that event knowledge and object-scene knowledge interacted in guiding fixations, with participants preferentially fixating objects with lower scores when event knowledge was irrelevant. Moreover, semantic similarity interacted with object size in the guidance of fixations. Our findings highlight the intricate relationship between event knowledge, object-scene knowledge and object size in guiding gaze