HAL-CentraleSupelec
Not a member yet
77624 research outputs found
Sort by
Reducing Aleatoric and Epistemic Uncertainty through Multi-modal Data Acquisition
No full textTo generate accurate and reliable predictions, modern AI systems need to combine data from multiple modalities, such as text, images, audio, spreadsheets, and time series. Multi-modal data introduces new opportunities and challenges for disentangling uncertainty: it is commonly assumed in the machine learning community that epistemic uncertainty can be reduced by collecting more data, while aleatoric uncertainty is irreducible. However, this assumption is challenged in modern AI systems when information is obtained from different modalities. This paper introduces an innovative data acquisition framework where uncertainty disentanglement leads to actionable decisions, allowing sampling in two directions: sample size and data modality. The main hypothesis is that aleatoric uncertainty decreases as the number of modalities increases, while epistemic uncertainty decreases by collecting more observations. We provide proof-of-concept implementations on two multi-modal datasets to showcase our data acquisition framework, which combines ideas from active learning, active feature acquisition and uncertainty quantification
High-Order Galois Automorphisms for TNFS Linear Algebra
No full textThe Number Field Sieve algorithm and its variants are the best known algorithms to solve the discrete logarithm problem in finite fields. When the extension degree is composite, the Tower variant TNFS is the most efficient. Looking at finite fields with composite extension degrees such as and is motivated by pairing-based cryptography that does not yet have a good quantum-resistant equivalent. The two most costly steps in TNFS are the relation collection} and linear algebra steps. Although the use of order Galois automorphisms allows one to accelerate the relation collection step by a factor of , their use to accelerate the linear algebra step remains an open problem. In previous work, this problem is solved for , leveraging a quadratic acceleration factor equal to .In this article, we bring a solution both for and . We propose a new construction that allows the use of an order (resp. ) Galois automorphism in any finite field (resp. ), thus accelerating the linear algebra step with approximately a factor of (resp. ). Moreover, we provide a SageMath implementation of TNFS and our construction, and validate our findings on small examples
Exploring Surface Seamless Mappings Toward Hexahedral Meshing
No full textMeshes provide an efficient and compact method for representing surfaces and volumes in computational settings, enabling numerical simulations of physical processes. Among the various mesh structures, \emph{quadrilateral and hexahedral meshes} are particularly attractive owing to their structured, grid-like nature. This regularity, interrupted only at a small number of singularities, yields significant benefits in terms of computational speed, accuracy, and parallelization efficiency.However, generating structured meshes remains a major challenge. This manuscript investigates mesh generation using the notion of \emph{seamless mappings}, which align surfaces or volumes to a regular grid. The primary goal of this study is to clarify the fundamental differences between quadrilateral (2D) and hexahedral (3D) mesh generation by systematically extending 2D algorithms to a 3D setting.The first method explored is a direct approach: solving the equations that define seamless maps. Although this strategy is tractable in two dimensions, it fails to converge in three dimensions. The root cause is identified as a conflict between enforcing seamlessness and local injectivity of the map.The second approach is based on a formulation using Cartan's moving frames, which allows greater flexibility in the placement of singularities. While this model improves the results in the 2D case, it still does not yield valid mappings in 3D, suggesting deeper structural challenges in higher dimensions.To address these limitations, a third method introduces a new deformation model: \emph{rectangular parametrizations}. These are defined as mappings that send infinitesimal, axis-aligned squares in a plane to rectangles on a surface. Crucially, injectivity is incorporated into the formulation. This model is robust and practical for surfaces. %, showing strong potential for applications such as fluid simulation.Extending this approach to 3D presents several open problems, including the existence of such mappings in volumetric domains and the difficulty in automatically placing valid singularities. These challenges and future research directions are explored in the final chapter.Les maillages offrent un moyen efficace et compact de représenter des surfaces et des volumes pour les simulations numériques de processus physiques. Parmi les différentes structures de maillage, les \emph{maillages quadrilatéraux et hexaédriques} sont particulièrement intéressants en raison de leur nature structurée, semblable à une grille. Cette régularité, seulement interrompue par un petit nombre de singularités, apporte des avantages significatifs en termes de vitesse de calcul, de précision et d'efficacité de parallélisation.Cependant, la génération de tels maillages structurés demeure un défi majeur. Ce manuscrit explore la génération de maillages à l'aide de la notion de \emph{carte sans couture} (\textit{seamless mappings}), qui alignent une surface ou un volume sur une grille régulière. L'un des objectifs de ce travail est de clarifier les différences fondamentales entre la génération de maillages quadrilatéraux (2D) et hexaédriques (3D), en généralisant les algorithmes 2D au cadre tridimensionnel.La première méthode étudiée repose sur une approche directe : résoudre les équations qui définissent les cartes sans couture. Si cette stratégie est praticable en deux dimensions, elle échoue à converger en trois dimensions. La cause profonde identifiée réside dans un conflit entre l'exigence de continuité et l'injectivité locale de la carte.La deuxième approche s'appuie sur une formulation utilisant les repères mobiles de Cartan, offrant une plus grande flexibilité dans le placement des singularités. Ce modèle améliore les résultats en 2D, mais ne permet toujours pas d'obtenir des cartes valides en 3D, ce qui suggère des obstacles structurels plus profonds dans les dimensions supérieures.Pour contourner ces limitations, une troisième méthode introduit un nouveau modèle de déformation : les \emph{paramétrisations rectangulaires}. Celles-ci sont définies comme des cartes envoyant des carrés infinitésimaux, alignés sur les axes du plan, vers des rectangles sur la surface. L'injectivité est intégrée directement dans la formulation. Sur les surfaces, ce modèle se révèle robuste et pratique.L'extension de cette approche au cas 3D soulève plusieurs problèmes ouverts, notamment l'existence de tels cartes dans des domaines volumiques et la difficulté de placer automatiquement des singularités valides. Ces défis, ainsi que des pistes de recherche futures, sont abordés dans le dernier chapitre
FDG-VTP: A Fully Decentralized Gossip Vehicular Trajectory Prediction Model
No full textInternational audienceVehicular ad hoc networks (VANETs) play a crucial role in modern intelligent transportation systems by enabling real-time communication between vehicles. To enhance vehicle trajectory prediction in VANETs, increasingly advanced and precise methods are being developed, aiming to improve traffic efficiency and road safety. Robust, resilient, and low-cost communication algorithms are also being proposed. However, these methods focus solely on the accuracy of the predictions, reducing communication costs, or improving convergence. In this study, we propose a method that provides both highly accurate trajectory prediction and an efficient fully decentralized communication algorithm with very low overhead, rapid convergence, and low computation costs. The proposed prediction approach is based on the Transformer model adapted to enhance trajectory prediction accuracy, while reducing computation time and communication costs. For the communication algorithm, the gossip learning approach is adopted using an aggregation algorithm based on the reliability of the received models and a communication logic based on communication frequencies evolving according to the progress of local training iterations within the target vehicle. Evaluations on the NGSIM US 101 and US I80 datasets demonstrate that the proposed method outperforms the comparison models in terms of higher prediction accuracy, while also ensuring lower communication costs and computation times compared to both centralized and fully decentralized state-ofthe-art models, all within a fully decentralized environment
Formalizing Representation Theorems for a Logical Framework with Rewriting
No full textRepresentation theorems for formal systems often take the form of an inductive translation that satisfies certain invariants, which are proved inductively. Theory morphisms and logical relations are common patterns of such inductive constructions. They allow representing the translation and the proofs of the invariants as a set of translation rules, corresponding to the cases of the inductions. Importantly, establishing the invariants is reduced to checking a finite set of, typically decidable, statements. Therefore, in a framework supporting theory morphisms and logical relations, translations that fit one of these patterns become much easier to formalize and to verify.The lambdaPi-calculus modulo rewriting is a logical framework designed for representing and translating between formal systems that has previously not systematically supported such patterns. In this paper, we extend it with theory morphisms and logical relations. We apply these to define and verify invariants for a number of translations between formal systems. In doing so, we identify some best practices that enable us to obtain elegant novel formalizations of some challenging translations, in particular sort-erasure translations from sorted to unsorted languages
Avatar Standardization Efforts for Interoperable Metaverse Services: Towards a Seamless Avatar-As-a-Service Ecosystem
No full textInternational audienceThe metaverse has emerged as a transformative platform that blends physical and digital realities, supporting immersive communication, interaction, and service delivery. At the core of these experiences are avatars, digital embodiments of users, that serve as interactive interfaces across a wide range of metaverse services, including virtual collaboration, education, healthcare, entertainment, e-commerce, etc. However, the fragmented ecosystem of tools, platforms, and proprietary formats has led to significant challenges in avatar interoperability, preventing seamless user identity and experience across applications. In response, numerous international standardization bodies have launched initiatives to define interoperable and modular avatar frameworks. This survey represents the first comprehensive review of these initiatives with a focus on service-oriented requirements, revealing the concept of Avatar-as-a-Service (AaaS). It examines both historical foundations and recent advances, with a particular attention given to the emerging ISO/IEC 23090-39 standard (MPEG-I Part 39 Avatar Representation Format, ARF). By synthesizing technical approaches, standardization efforts, and open challenges, this work presents a roadmap for the evolution of AaaS and highlights emerging opportunities for delivering personalized, high-fidelity, and cross-compatible avatar experiences to support the next generation of metaverse services
PyroBuildS: Speeding up the exploration of large configuration spaces with incremental build
No full textInternational audienceSoftware developers are acutely aware that software build is an essential but resource-intensive step in any software development process, all the more when building large and/or highly configurable systems, whose vast number of configuration options leads to an explosion in the number of variants to build and evaluate. A potential approach to speed up the builds of multiple configurations is to do incremental build, i.e., to not clean the build environment and reuse previous builds when building a new configuration. Previous exploratory studies showed some benefits and limitations of incremental build, but mainly on small configurable software systems and on a limited set of close configurations. However, for large configuration spaces, little is known whether the large distance across configurations impacts the correctness and efficiency of incremental build.This paper presents PyroBuildS, a new approach to speed up incremental builds while keeping reproducibility, featuring a configuration variation operator parameterized by two deny lists of problematic options and a mutation size (diversity).We evaluate PyroBuildS through an empirical study on three large complex configurable systems, namely Linux, BusyBox, and ToyBox, with respectively 18637, 1078, 330 configuration options. We first show that for all configurations PyroBuildS produces the exact same binaries as a clean build, except for Linux with some non-reproducible random configurations. We identify the reasons why incremental build speeds up or slows down the build of large configuration spaces – a knowledge that can be integrated into PyroBuildS. Incremental build systematically pays off, since problematic options are avoided in the first place — something only PyroBuildS does. We also show that a naive use of incremental build on random Linux configurations backfires, taking more time than clean builds. Thus, PyroBuildS controls diversity to avoid too many differences across configurations to perform efficient incremental builds.Thanks to its ability to operate over non-problematic options and close enough configurations, PyroBuildS significantly speeds up the exploration of large configuration spaces, with a gain in build time from 16% to 22% in all three systems with mutated configurations. Finally, with random configurations, PyroBuildS also speeds up the build time from 15% to 20% for ToyBox and BusyBox
Shallow‐Donor‐Controlled Conduction in ZnMgO: Mg‐Induced Transition Between Variable‐Range and Nearest‐Neighbor Hopping Mechanisms
No full textInternational audienceZnMgO thin films (0%–3% Mg) were deposited via ultrasonic spray pyrolysis and characterized from 40 to 320 K. Three conduction regimes were identified: Mott variable‐range hopping (VRH), nearest‐neighbor hopping (NNH), and thermally activated conduction. At 0% Mg, VRH dominates; at 1–2% Mg, VRH and NNH coexist; and at 3% Mg, NNH prevails. Key parameters include a density of states near the Fermi level of 5.7 × 10^(19) cm^(−3) eV^(−1) (0% Mg) and 3.9 × 10^(18) cm^(−3) eV^(−1) (1% Mg), localization lengths of 11.0 nm (0% Mg) and 23.4 nm (1% Mg), hopping energies of 3.1 meV (VRH, 0% Mg) and 4.8 meV (VRH, 1% Mg), and NNH activation energies from 5.5 meV (1% Mg) to 9.9 meV (3% Mg). High‐temperature activation energy increases from 26.9 to 70.2 meV. Optical measurements show bandgap widening from 3.264 to 3.308 eV, aligning with transport trends. These results advance ZnMgO conduction understanding for optoelectronic and sensing applications
Shift your Focus for the Greater Good: Improving Fairness at no cost for Accuracy and Diversity in News Recommender Systems
No full textInternational audienceIn today’s digital landscape, recommender systems assist users in navigating the vast amount of available data. Within the realm of information access, a subset of such systems called News Recommender Systems help users find news content that interests them. However, by prioritizing traditional accuracy-focused optimization, these systems contribute to the formation of filter bubbles, restricting users’ exposure to diverse viewpoints and exacerbating polarization. To address this issue, beyond-accuracy factors like diversity have been integrated into recommendation. Yet, such approaches can be ineffective and even unintentionally influence user opinions. This raises major ethical concerns as systems lack the legitimacy to shape opinions. This paper presents the ADF framework, a novel approach designed to optimize accuracy, diversity, and fairness simultaneously. Unlike conventional models that manage fairness in a trade-off, ADF establishes fairness as a core constraint. The framework relies on an innovative fairness-constrained diversification strategy, ensuring that users are exposed to a broader range of opinions, without being oriented towards specific viewpoints. ADF is adaptable to various diversity metrics and provides personalized diversification, independent of the underlying recommendation algorithms. Through real-world benchmark datasets and multiple recommendation models, experimental evaluation confirmed that ADF limits the impact on accuracy, enhances diversity, and crucially upholds fairness
Journée d'étude "Les questions et le discours" : Avant-propos
No full textInternational audienceThe workshop "Questions and the Discourse" aims to celebrate Valentin D. Richard's defense and gather colleagues working on topics related to his PhD dissertation. This presentation introduces the talks of the three speakers and connections between their work and Richard's thesis.The unity behind the talks rely on the pervasive presence of alternatives in natural language structures and models. Alternatives construed as Questions under Discussions can help us keep track of conversation topics. Regarding indefinites, modeling their interpretation with sets of possible values allows us to capture variation and dependence between individuals. Finally, modeling disjunction with truthmakers provides a finer model for certain subtle effects.La journée d'étude « Questions and the Discourse » (Les questions et le discours) célèbre la soutenance de thèse de Valentin D. Richard en réunissant des collègues travaillant sur des sujets liés à sa thèse de doctorat. Cette présentation est une introduction aux exposés des trois intervenants et aux liens entre leurs travaux et le manuscrit de Richard.L'unité derrière ces trois exposés réside dans la présence omniprésente d'alternatives dans les structures et les modèles de la langue. Les alternatives interprétées comme des "questions en discussion" (Questions under Discussion) sont utiles pour suivre la trace des sujets de conversation. En ce qui concerne les indéfinis, modéliser leur interprétation à l'aide d'ensembles de valeurs possibles nous permet de représenter les variations et les dépendances entre les individus. Enfin, modéliser la disjonction à l'aide de vérifieurs exacts (truthmakers) permet d'obtenir un modèle plus précis dans le cas de certains effets subtils