Archive ouverte de Centrale Lyon
Not a member yet
32420 research outputs found
Sort by
Towards Better Interpretability of Sepsis Prediction by Deep Neural Networks with Variable-wise Attribution Maps
No full textInternational audienceBecause of the multi-symptomatic nature of sepsis, its prediction is challenging as it requires considering subtle changes in multiple monitored variables across time. Recent works based on deep neural networks improved the prediction performance, but still suffer from poor interpretability. Critical for healthcare applications, we propose to improve this aspect by separating time variables in the convolution layers of a sepsis prediction network. We reveal the improvement in interpretability capacity with the use of gradient-based attributions on high-level intermediate features and through a metric correlating the variable attribution with the prediction for perturbed pathologic samples. With 171,945 patients from the MIMIC-IV database, we demonstrate that our method not only maintains classification performances at similar network parameters count, but also substantially improves the faithfulness of per-variable attributions
ImFace++: A Sophisticated Nonlinear 3D Morphable Face Model With Implicit Neural Representations
No full textInternational audienceAccurate representations of 3D faces are of paramount importance in various computer vision and graphics applications. However, the challenges persist due to the limitations imposed by data discretization and model linearity, which hinder the precise capture of identity and expression clues in current studies. This paper presents a novel 3D morphable face model, named ImFace++, to learn a sophisticated and continuous space with implicit neural representations. ImFace++ first constructs two explicitly disentangled deformation fields to model complex shapes associated with identities and expressions, respectively, which simultaneously facilitate automatic learning of point-to-point correspondences across diverse facial shapes. To capture more sophisticated facial details, a refinement displacement field within the template space is further incorporated, enabling fine-grained learning of individual-specific facial details. Furthermore, a Neural Blend-Field is designed to reinforce the representation capabilities through adaptive blending of an array of local fields. In addition to ImFace++, we devise an improved learning strategy to extend expression embeddings, allowing for a broader range of expression variations. Comprehensive qualitative and quantitative evaluation demonstrates that ImFace++ significantly advances the state-of-the-art in terms of both face reconstruction fidelity and correspondence accuracy
Novel Electrical Characterization Method for Antiferroelectrics using a Positive Up Negative Down Approach
No full textInternational audienceThis study demonstrates the effectiveness of AFE-PUND, a revisited Positive Up Negative Down (PUND) protocol for characterizing antiferroelectric (AFE) materials, in analyzing films across different thicknesses, revealing key trends. The proposed AFE-PUND method enables the isolation of switching currents from non-switching contributions, allowing precise extraction of remanent polarization and coercive field from hysteresis loops. The remanent polarization increases with film thickness, reflecting enhanced domain stability, while endurance cycles highlight the wake-up effect and its eventual degradation due to fatigue in thicker films. Similarly, coercive fields decrease with thickness, indicating reduced switching barriers and a clearer transition between tetragonal and orthorhombic phases. The method provides valuable insights into micro-structural influences, such as defect accumulation, grain size, and domain wall pinning, which critically affect device performance. AFE-PUND thus establishes itself as an essential tool for advancing the understanding and optimization of antiferroelectric materials
Numerical modeling of hydrogel scaffold anisotropy during extrusion-based 3D printing for tissue engineering
No full textInternational audienceExtrusion-based 3D printing is a widely utilized tool in tissue engineering, offering precise 3D control of bioinks to construct organ-sized biomaterial objects with hierarchically organized cellularized scaffolds. The internal organization of scaffold constituents must replicate the structural anisotropy of the targeted tissue to effectively promote cellular behavior during 3D cell culture. The choice of polymers in the bioink and extrusion process topological properties significantly impact tissue engineering constructs' structural anisotropy and cellular response. Our study employed a hydrogel bioink consisting of fibrinogen, alginate, and gelatin, providing biocompatibility, printability, and shape retention post-printing. Topological properties in flowing polymers are determined by macromolecule conformation, namely orientation and stretch degree. We utilized the micro-macro approach to describe hydrogel macromolecule orientation during extrusion, offering a two-scale fluid behavior description. The study aimed to use the Fokker-Planck equation to represent constituent population (polymer chain) state within a hydrogel's representative elementary volume during extrusion-based 3D printing. Our findings indicate that a high shear rate drives constituent orientation in tubular nozzle syringe setups, overcoming fluid rheological behavior. Additionally, the interaction coefficient (C_i), representing microscopic fluid particle interaction, surpasses hydrogel behavior for constituent orientation prediction. This approach provides an initial but robust framework to model scaffold anisotropy, enabling optimization of the extrusion process while maintaining computational feasibility
Stochastic optimization of a mass-in-mass cell with piecewise hybrid nonlinear-linear restoring force
No full textInternational audienceThis article investigates the optimization under uncertainty of a mass-in-mass meta-cell for its potential use within a metamaterial. The specificity of the proposed mass-in-mass system stems from the hybrid nonlinear-linear stiffness at the inner level. It is well known that these systems exhibit high sensitivity to small perturbations in loading conditions or design parameters. In fact, the sensitivity is such that the system can exhibit discontinuous behaviors. Therefore the proposed optimization approach not only accounts for sources of uncertainties but also can handle discontinuous responses. The objective of the stochastic optimization is to find the stiffness properties of the mass-in-mass system which minimize the expected value of a specific efficiency metric. In order to better understand the system’s dynamic behavior and the origins of the discontinuities, slow invariant manifolds and frequency response curves are provided. The efficiency of the optimized system with hybrid stiffness is compared with that of a similar optimized system featuring pure cubic nonlinearity
Entropy-Guided Self-Regulated Learning Without Forgetting for Distribution-Shift Continual Learning with blurred task boundaries
No full textInternational audienceContinual Learning (CL) aims to endow machines with the human-like ability to continuously acquire novel knowledge while retaining previously learned experiences. Recent research on CL has focused on Domain-Incremental Learning (DIL) or Class-Incremental Learning (CIL) with well-defined task boundaries. However, for real-life applications, e.g., waste sorting, robotic grasping, etc., the model needs to be constantly updated to fit new data. Additionally, there is usually an overlap between new and old data. Thus, task boundaries may not be well defined, and a more smooth scenario is needed. In this paper, we propose a more general scenario, namely Distribution-Shift Incremental Learning (DS-IL), which enables soft task boundaries with possible mixtures of data distributions over tasks and thereby subsumes the two previous CL scenarios: DIL and CIL are simply DS-IL. Moreover, given the increasingly greater importance of data privacy in real-life applications and, incidentally, data storage efficiency, we further introduce an entropy-guided self-regulated distillation process \textbf{without memory}, which leverages data similarities between tasks with soft-boundaries. Experimented on a variety of datasets, our proposed method outperforms or matches state-of-the-art continual learning methods
Periodicity and longtime diffusion for mean field systems in
No full text43 pagesInternational audienceWe study in this paper the longtime behavior of some large but finite populations of interacting stochastic differential equations whose (infinite population) limit Fokker-Planck PDE admits a stable periodic solution. We show that the empirical measure for the population of size stays close to the periodic solution, but with a random dephasing at the timescale that converges weakly to a Brownian motion with constant drift
Convertisseurs résonants multi-MHz : concepts, outils et méthodologies
No full textThe growing interest in high-frequency power conversion has driven significant research into resonant topologies operating in the multi-MHz range. While these converters offer compelling benefits in terms of size reduction, efficiency, and dynamic performance, their design presents unique challenges that extend beyond traditional power electronics paradigms. This thesis investigates the conceptual, methodological, and practical aspects of designing multi-MHz resonant converters. It focuses on the development and application of dedicated tools and design methodologies suited to this frequency domain. A comprehensive framework is proposed to support the design process, from early conceptual modeling to detailed implementation. The work adopts an exploratory approach and incorporates a historical perspective to contextualize the proposed methodologies within the broader evolution of resonant converter research. The result is a set of insights and analytical methods aimed at supporting the design of next-generation resonant converters in the multi-MHz regime.L’intérêt croissant pour la conversion d’énergie à haute fréquence a suscité de nombreux travaux de recherche sur les topologies résonantes fonctionnant dans la gamme de fréquences supérieures au MHz. Bien que ces convertisseurs présentent des avantages notables en termes de miniaturisation, de rendement et de performances dynamiques, leur conception soulève des défis spécifiques qui dépassent les cadres classiques de l’électronique de puissance. Cette thèse examine les aspects conceptuels, méthodologiques et pratiques de la conception de convertisseurs résonants multi-MHz. Elle se concentre sur le développement et l’application de modèles analytiques et de méthodologies de conception adaptées à ce domaine fréquentiel. Un cadre de travail complet est proposé pour accompagner le processus de conception, depuis la modélisation conceptuelle initiale jusqu’aux considération pratiques finales. Le travail adopte une approche exploratoire et intègre une perspective historique afin de replacer les méthodologies proposées dans le contexte plus large de la recherche sur les convertisseurs résonants. Il en résulte un ensemble d’analyses et de méthodes visant à soutenir la conception des convertisseurs multi-MHz résonants de nouvelle génération
Apprentissage collaboratif de confiance : Personnalisation, confidentialité et cobustesse en environnements décentralisés
No full textTwo decades ago, the emergence of Web 2.0 transformed our digital experiences by enabling users to create and share content through various platforms, leading to an ever-growing volume of user-generated data. Coupled with recent advancements in Machine Learning, this abundance of data has enabled the emergence of a wide range of learning-based services (e.g., AI chatbots, voice assistants, fraud detection). Traditionally, training the models behind these services involved collecting user data through centralized entities—often the service providers themselves. However, this data is often sensitive in nature (e.g., health records, mobility patterns), as illustrated by the numerous privacy scandals that unfolded over the past decade. These incidents have raised user awareness around privacy and motivated the introduction of privacy regulations such as the General Data Protection Regulation (GDPR). In this context, rethinking classical learning paradigms and designing privacy-aware alternatives has become a critical need. Federated Learning (FL) emerged in 2017 as a new paradigm that promises privacy and governance by allowing users to collaboratively train a machine learning model without ever sharing their raw data. This gave rise to a broader class of algorithms known as Collaborative Learning. It is now well-established that FL, in its original form, suffers from several drawbacks—the most significant being its reliance on a central server, which represents a single point of failure, and its vulnerability to various privacy and robustness attacks. It is in this landscape that Gossip Learning (GL) appeared, with the promise of full decentralization and several appealing properties inherited from gossip communication protocols. In this paradigm, each node behaves somewhat like an FL server, coordinating learning with its own neighbors over a decentralized, time-dependent communication graph. Yet, it remains unclear whether the decentralized and randomized nature of GL inherently enables it to overcome the limitations of FL, particularly in terms of personalization (i.e., generating satisfactory models for individuals), resilience to privacy attacks, and robustness against adversarial behavior. This thesis aims to shed light on these open questions. First, we compare the performance of individual models in FL and GL, before introducing PEPPER, a Gossip Learning framework that harnesses GL's full potential and outperforms FL in personalized settings. Second, we propose CIA, a Community Inference Attack, which we use to audit the privacy vulnerabilities of both FL and GL—leading to the insight that GL demonstrates an inherent resilience to CIA and comparison-based inference attacks. Finally, we study the vulnerability of GL to robustness attacks (i.e., model poisoning). While we find that GL's dynamic nature can make it more vulnerable in some cases, we propose GRANITE, a framework for robust Gossip Learning over dynamic graphs.Overall, our work shows that while GL is still in its early stages, it holds the potential to become a reliable and more suitable paradigm for Collaborative Learning.Il y a une vingtaine d'années, l'émergence du Web 2.0 a profondément transformé notre rapport au numérique en permettant aux utilisateurs de créer et partager du contenu sur une multitude de plateformes, générant ainsi un volume croissant de données. Combinée aux récents progrès en apprentissage automatique, cette abondance de données a permis l'apparition de nombreux services basés sur l'apprentissage (e.g., agents conversationnels, assistants vocaux, détection de fraude). Traditionnellement, l'entraînement des modèles sous-jacents à ces services reposait sur la collecte de données sensibles par des entités centralisées, souvent les fournisseurs de service eux-mêmes. Cette centralisation a soulevé de sérieuses préoccupations en matière de confidentialité, comme en témoignent les multiples scandales liés à la vie privée au cours de la dernière décennie. En réponse, plusieurs réglementations, telles que le Règlement Général sur la Protection des Données (RGPD), ont été mises en place. Dans ce contexte, repenser les paradigmes d'apprentissage classiques pour les rendre plus respectueux de la vie privée est devenu essentiel. L'apprentissage fédéré a vu le jour en 2017 avec la promesse de préserver la confidentialité des données en permettant à des utilisateurs d'entraîner collaborativement un modèle sans jamais partager leurs données brutes. Ce paradigme a donné naissance à une classe plus large d'approches appelée apprentissage collaboratif. Il est désormais bien établi que l'apprentissage fédéré souffre de plusieurs limitations, notamment sa dépendance à un serveur central (point de défaillance unique) et sa vulnérabilité à diverses attaques en confidentialité et robustesse. C'est dans cette optique que l'apprentissage par commérage a été proposé, avec la promesse d'une décentralisation totale, s'appuyant sur des protocoles de communication pair-à-pair, dits protocoles de bavardage. Dans ce paradigme, chaque nœud agit comme un mini-serveur fédéré, coordonnant l'apprentissage avec ses voisins via un graphe de communication dynamique. Toutefois, il reste à démontrer si cette approche permet réellement de dépasser les limites de l'apprentissage fédéré, notamment en matière de personnalisation, de protection de la vie privée et de robustesse face aux comportements malveillants. Cette thèse s'attache à explorer ces questions. Dans un premier temps, nous comparons les performances individuelles des modèles produits par ces deux paradigmes, avant de proposer PEPPER, une structure logicielle permettant d'exploiter pleinement le potentiel de l'apprentissage par commérage à des fins de personnalisation, et de surpasser l'approche fédérée sur cet aspect. Dans un second temps, nous introduisons CIA, une attaque d'inférence de communautés utilisée pour auditer les vulnérabilités de confidentialité des deux paradigmes. Cette étude révèle une certaine résilience intrinsèque de l'apprentissage par commérage face à ce type d'attaques d'inférence comparative. Enfin, nous nous penchons sur la robustesse de ce paradigme face aux attaques par empoisonnement de modèles. Bien que sa nature dynamique puisse l'exposer davantage à ce type de menaces, nous proposons GRANITE, un cadre logiciel robuste pour l'apprentissage par commérage sur des graphes dynamiques. Dans l'ensemble, ce travail met en évidence le potentiel de l'apprentissage par commérage à s'imposer comme une alternative crédible à long terme pour des systèmes d'apprentissage décentralisés, transparents et centrés sur l'utilisateur
Fast Spectral Data Simulation from sRGB Images Using K-Dimensional Search Trees
No full textInternational audienceAbstract Spectral imaging is a powerful tool to analyse and manipulate the spectral content of a scene in a spatial context. Nevertheless, it is a complex and expensive tool, and limited under certain conditions (non-ambient temperature conditions, low light etc.). This study introduces a fast and efficient alternative to generate plausible spectral data from the CIELAB values of an sRGB image and a spectral reflectance database of +7 million colours. The proposed methodology relies on the compartmentalization capabilities of K-Dimension search trees to quickly identify, under a chosen light source, the spectral reflectance sample, that most closely matches the La*b* values of each test pixel. On average, the algorithm takes 6 s to substitute RGB values with plausible spectral data for full size high-definition images. The developed technique allows lighting or imaging researchers to rapidly generate spectral data directly from RGB images. The simulated spectral data, when, converted to CIELAB values, produces an image perceptually identical to the original LAB values for 95% pixels (<2 Delta Eab*). Moreover, by simulating spectral data under actual light source, researchers can achieve high spectral accuracy comparable to actual spectral data