Portail HAL de l'Université du Littoral Côte d'Opale
Not a member yet
11255 research outputs found
Sort by
On the Robustness of BFGS to Positive and Negative Noise Outliers on the BBOB Test Suite
No full textInternational audienceWe investigate the impact of outlier noise on the performance of the scipy.optimize implementation of the quasi-Newton BFGS solver. Using the BBOB testbed corrupted with positive—making solution look worse than what they are—or negative—making solutions look better than what they are—outliers simulated with a Cauchy distribution with a probability p, we analyze how the performance is impacted. We show that the impact of positive or negative noise outliers is almost symmetric, that on simple problems BFGS has some robustness to noise, and that for ill-conditionned problems BFGS appears to fail when p or the dimension is too high
Adaptive Unsupervised Graph Convolution Network for Data Clustering with Graph Reconstruction
No full textInternational audienceIn recent years, graph clustering has emerged as one of the most challenging problems in the field of deep learning. With the increasing complexity of real-world networks, such as social and biological networks, more and more effective methods are needed to organize and understand these structures. Various cognition-based techniques have been explored for classifying nodes within these graphs, and graph convolution networks (GCNs) have attracted great interest. GCNs, a deep semi-supervised learning approach, provide a powerful framework for learning node representations by utilizing both local and global graph information. By iteratively aggregating information from neighboring nodes, GCNs effectively capture the intricate relationships and dependencies within complex networks. We introduce a novel deep unsupervised learning scheme built upon the foundation of GCN architecture. The key contributions are outlined below. First, the entire architecture is trained with three unsupervised learning losses. The first loss focuses on kernelized features that use node attributes to reflect the information extracted from the data. The second loss leverages spectral smoothness that uses connections between nodes to capture global cluster structure. The third loss is based on graph reconstruction that introduces additional regularization of the representation of nodes by the output of the model. Second, the spectral smoothing loss involves an adaptive approach using an additional graph matrix associated with the node representations. This adaptive integration of additional structural information increases the learning efficiency during the training phase. The adaptive fused graph used for spectral smoothing loss incorporates structural insights derived from both data features and deep node representations. To assess the performance of our approach, we conducted extensive experimental evaluations on four benchmark datasets widely used in the field of graph clustering. These datasets were carefully selected to cover diverse domains and varying degrees of complexity, ensuring a comprehensive evaluation of our method’s efficacy. Our results showcase the remarkable performance of our unsupervised GCN across multiple metrics, surpassing other state-of-the-art graph neural network techniques in terms of clustering accuracy, purity, and other relevant measures. Notably, our method consistently outperforms competing approaches across the used datasets, demonstrating its versatility and effectiveness in handling various real-world scenarios
Le lecteur : une figure toute-puissante ? Rôle du lecteur et échec d’une lecture programmée dans l’œuvre d’Emmanuel Carrère
No full textInternational audienceSince the publication of L’Amie du jaguar in 1984, Emmanuel Carrère has positioned the reader at the core of his literary enterprise, at times blurring the line between reader and “fictional character” (Hétu 1988: 58). A playful complicity emerges between writer and reader, fostering a dynamic of engagement. Carrère’s poetics encourage the reader to take an active role in the construction of meaning by granting a space for textual freedom. Yet this autonomy is often more apparent than real, as the author ultimately seeks to retain absolute control over the narrative. This article examines the strategies Carrère employs to seduce, dominate, and manipulate the reader, while also considering the forms of resistance the reader may offer in response to such authorial authority.Depuis L’Amie du jaguar publié en 1984, Emmanuel Carrère accorde une place centrale au lecteur, jusqu’à le confondre avec un « personnage de fiction » (Hétu 1988 : 58). Un rapport de complicité, proche du jeu, s’instaure entre l’écrivain et son destinataire. À travers sa poétique, Carrère invite le lecteur à s’impliquer activement dans la construction du sens, en lui octroyant un espace de liberté textuelle. Cependant, cette autonomie demeure souvent illusoire, l’auteur aspirant à demeurer maître absolu du récit. Cette contribution explore les stratégies mises en œuvre par l’auteur pour séduire, contrôler et soumettre le lecteur tout en interrogeant les résistances que celui-ci peut opposer face à cette domination
La mobilisation des employeurs pour les salariés victimes de violences domestiques
No full textInternational audienc
Urban projects in France: navigating neoliberalism, justice, and commons planning
No full textInternational audienceThis review paper examines the concept of urban projects within the evolving landscape of urban planning theory and practice, with a particular focus on the French context. Through a targeted review of French and English language literature, we trace the development of planning models from rational comprehensive approaches to more collaborative and strategic models. We argue that the urban project emerges as a multifaceted construct, embodying processes, outcomes, and actor-rules dynamics. The paper explores three key research subjects in contemporary urban projects: the interplay between neoliberal and the just city paradigms, the divergence between judicial and spatial understandings of justice, and the potential of commons and social innovation as an emerging paradigm and a response to neoliberal paradigm and spatial inequalities. This research synthesises existing knowledge and identifies key areas for future investigation, offering insights that can inform both theoretical discourse and practical applications in urban planning. Article Highlights• The Urban Project is a France's strategic planning instrument, not inherently aligned with neoliberal agendas • The "Just City" paradigm fails to prevent neoliberal outcomes; community-based innovation offers key insights to address spatial injustices and inequalities • Judicial and spatial justice interpretations vary, potentially leading to spatial injustices and inequalities Keywords Urban projects • Just city • Neoliberal paradigm • Commons planning • French planning theory * Youness Achmani,</div
Enzymatic-controlled production and characterization of bio-based oligogalacturonides for application in wheat protection against powdery mildew.
No full textInternational audienc
Hybridation de Modèles d’IA avec des Classifieurs Ontologiquement Explicables
No full textInternational audienceThe research in eXplainable Artificial Intelligence (XAI) has emphasized the need to create models based on domain knowledge to make them explainable from their users’ perspective. A significant portion of current work focuses on designing AI models that integrate the domain experts’ semantics and way of reasoning. Drawing inspiration from Concept Based Models and Neuro-Symbolic AI, we propose a hybrid architecture for constructing AI pipelines that utilize Machine Learning to extract domain concepts and symbolic reasoning to predict an explainable classification. The core of this proposal is the OntoClassifier, a module that uses domain ontologies to automatically generate ontologically explainable classifiers. We describe the proposed approach and architecture, detailing the implementation and capabilities of the OntoClassifier. The solution is applied in Computer Vision and is illustrated using the Pizzaïolo Dataset.La recherche en Intelligence Artificielle Explicable (XAI) a souligné la nécessité de créer des modèles basés sur les connaissances du domaine pour qu'ils soient explicables du point de vue de leurs utilisateurs. Une part importante des travaux actuels se concentre sur la conception de modèles d'IA qui intègrent la sémantique et le mode de raisonnement des experts du domaine. S'inspirant des Modèles Basés sur les Concepts et des approches Neuro-Symboliques, nous proposons une architecture hybride pour construire des pipelines d'IA qui utilisent l'Apprentissage Automatique pour extraire les concepts du domaine et un raisonnement symbolique pour prédire une classification explicable. Le coeur de cette proposition est l'OntoClassifier, un module qui utilise des ontologies de domaine pour générer automatiquement des classifieurs ontologiquement explicables. Nous décrivons l'approche et l'architecture proposées en détaillant l'implémentation et les capacités de l'OntoClassifier. La solution est appliquée en Vision par Ordinateur et est illustrée à l'aide du Pizzaïolo Dataset
Rigidity-Driven Structural Isomers in the NaCl–Ga2S3 System: Implications for Energy Storage
No full textInternational audienceAlternative energy sources require the search for innovative materials with promising functionalities. Systems with unusual chemical properties represent an insufficiently explored domain, concealing unexpected features. Using diffraction and Raman spectroscopy over a wide temperature range, supported by first-principles simulations, a rare phenomenon is unveiled: phase-dependent chemical interactions between binary components in the NaCl–Ga2S3 system. In this unique occurrence, previously intact binary crystalline species transform upon melting into mixed liquid structural isomers, forming bonds with new partners. The chemical combinatorics appears to be fully reversible for stable crystals and liquids. Despite this, rapidly frozen glasses out of thermodynamic equilibrium remain in a metastable isomeric state, offering remarkable properties, particularly a high room-temperature Na+ conductivity, comparable to the best sodium halide superionic conductors and therefore encouraging for sodium solid-state batteries and energy applications. A rigidity paradigm is responsible for the observed phenomenon, as the extremely constrained Ga2S3 crystal lattice does not survive viscous flow, breaking up at a short-range level. The removal of rigidity constraints and dense packing leads to a significant increase in empty space, which is the origin of high sodium diffusivity. Broadly, the rigidity-driven structural isomerism opens up an inspiring path to the discovery of atypical materials
Problèmes inverses de bathymétrie pour les ondes de surface et analyse spectrale des discrétisations isogéométriques de Laplace
No full textKnowledge of underwater topography (bathymetry) in rivers and oceans is indispens-able for many real-world applications, including the design of offshore platforms andharbors, marine navigation safety, and the prediction of tsunami propagation. Accurateestimation of this geometry is challenging and often requires in situ measurements,such as the use of acoustic sounding (sonar). However, these techniques are time-consuming and expensive, particularly for large aquatic domains. In this thesis, weapproach the problem from a modeling perspective, treating the bottom profile as avariable coefficient in the governing flow equations. Specifically, we model the flowdynamics with the general water waves system and then its shallow-water asymptoticsimplification (Saint-Venant system). For each model, we study the inverse problem ofreconstructing bathymetry from measurements taken at the free surface of the fluid(waves). Our methodology is to first establish the mathematical uniqueness and stabil-ity for the two inverse problems; then, develop fast and highly accurate solvers for thedetection in real-world settings.Because of this clear intersection of mathematics, numerics, and physics, we focus ontwo main axes: the analysis of these inverse problems and the Isogeometric Anal-ysis (IGA) finite-element discretization of second-order elliptic systems, such as thepotential-flow Laplace system associated with the first inverse problem. The thesis isstructured into four chapters. In the first chapter, we consider the (d + 1)-dimensionalgeneral water waves system and establish identifiability and log–log (and log) stability.Motivated by the ill-posed nature of this problem, the second chapter reformulatesthe inverse problem using the one-dimensional shallow water model and establishesuniqueness and Lipschitz stability. This chapter also proposes an efficient algorithm and provides numerical validation of the approach. The final two chapters concernIGA discretizations of second-order elliptic operators. The first presents a fast ana-lytic solver for Laplace systems on d-box domains, while the last chapter deals with aparticular case of a second-order elliptic operator with variable coefficients and offersinsights and new estimates for the eigenfrequencies of the IGA-discretized operator.La connaissance de la topographie sous-marine (bathymétrie) dans les rivières et les océans est indispensable pour de nombreuses applications réelles, notamment la conception de plateformes offshore et de ports, la sécurité de la navigation maritime, ainsi que la modélisation et prédiction de la propagation des tsunamis. L’estimation précise de cette géométrie est difficile et nécessite souvent des mesures sur site, par exemple au moyen de sondages acoustiques (sonar). Cependant, ces techniques sont longues et coûteuses, en particulier pour de larges zones maritimes ou fluviales. Dans cette thèse,nous adoptons une approche de modélisation, en considérant le profil du fond comme un coefficient variable dans les équations gouvernant l’écoulement. Plus précisément,nous modélisons la dynamique de l’écoulement avec le système général des ondes de surface, puis avec sa simplification asymptotique en eaux peu profondes (système de Saint-Venant). Pour chacun des deux modèles, nous étudions le problème inverse de reconstruction de la bathymétrie à partir de mesures effectuées à la surface libre du fluide (ondes). Notre méthodologie consiste d’abord à établir l’unicité et la stabilité mathématiques pour les deux problèmes inverses ; puis à développer des solveurs rapides et précis pour la détection dans des contextes réels.En raison de cette intersection claire entre les mathématiques, le calcul numérique et la physique, notre travail se structure autour de deux axes : l’analyse de ces problèmes inverses et la discrétisation par éléments finis basés sur l’analyse isogéométrique de systèmes elliptiques du second ordre, en particulier le système de Laplace modélisant l’écoulement potentiel, relié directement au premier problème inverse. La thèse est structurée en quatre chapitres. Dans le premier chapitre, nous considérons le système général des ondes de surface en dimension d + 1, avec d = 1, 2, et établissons l’identifiabilité et une stabilité de type log–log (et log). Motivé par la nature mal posée de ce problème, le deuxième chapitre reformule le problème inverse via le modèle uni-dimensionnel des eaux peu profondes et démontre l’unicité ainsi qu’une stabilité lips-chitzienne. Ce chapitre propose également un algorithme efficace accompagné d’une validation numérique de l’approche. Les deux derniers chapitres de la thèse concernent les discrétisations IGA d’opérateurs elliptiques du second ordre. Le premier présente un solveur analytique rapide pour des systèmes de Laplace sur des boîtes de dimension générale d, tandis que le dernier chapitre examine un cas particulier d’opérateur elliptique du second ordre à coefficients variables et fournit de nouvelles informations et estimations pour les fréquences propres de l’opérateur discrétisé par IGA