HAL Portal UTC Université de Technologie de Compiègne
Not a member yet
11652 research outputs found
Sort by
Traduction Agile et Novlangue Managériale: Ou comment se traduisent mal les concepts « américains »
No full textInternational audienc
Coupling of experiments and DEM simulation for powder flow understanding and prediction
No full textInternational audienc
Footprints of Data in a Classifier: Understanding the Privacy Risks and Solution Strategies
No full textInternational audienceThe widespread deployment of Artificial Intelligence (AI) across government and private industries brings both advancements and heightened privacy and security concerns. Article 17 of the General Data Protection Regulation (GDPR) mandates the Right to Erasure, requiring data to be permanently removed from a system to prevent potential compromise. While existing research primarily focuses on erasing sensitive data attributes, several passive data compromise mechanisms remain underexplored and unaddressed. One such issue arises from the residual footprints of training data embedded within predictive models. Performance disparities between test and training data can inadvertently reveal which data points were part of the training set, posing a privacy risk. This study examines how two fundamental aspects of classifier systems—training data quality and classifier training methodology—contribute to privacy vulnerabilities. Our theoretical analysis demonstrates that classifiers exhibit universal vulnerability under conditions of data imbalance and distributional shifts. Empirical findings reinforce our theoretical results, highlighting the significant role of training data quality in classifier susceptibility. Additionally, our study reveals that a classifier’s operational mechanism and architectural design impact its vulnerability. We further investigate mitigation strategies through data obfuscation techniques and analyze their impact on both privacy and classification performance. To aid practitioners, we introduce a privacy-performance trade-off index, providing a structured approach to balancing privacy protection with model effectiveness. The findings offer valuable insights for selecting classifiers and curating training data in diverse real-world applications
Comparison between different electrohysterogram data augmentation methods for better prediction of the delivery term
No full textInternational audienceThe early delivery of the infants can risk their lives andcause them serious health issues in the future. Artificial intelligencemodels that are trained on the uterine muscular contraction signals(electrohysterogram) have been shown to be efficient for predictingthe delivery term and hence decreasing its threat. However, the lackand imbalance of data lead to a biased and ineffective learningprocess of the employed AI model. In this manner, this paperinvestigates the possibility of data augmentation (DA) using the high-order message passing scheme used in the hypergraph neuralnetwork (HGNN). To prove the validity of the method, it wascompared to several DA methods, including synthetic minorityoversampling technique (SMOTE), noise addition, autoregressive(AR) modeling, and deep convolutional generative adversarialnetwork (DCGAN). The prediction F1-score values achieved by theemployed HGNN upon the application of the different DA techniqueswere 84.4 ± 12.6% (hypergraph-based), 82.5 ± 14.6% (SMOTE), 82.6± 15.8% (noise addition), and 77.9 ± 15.5% (AR modeling). The smallsize of our database was not enough to efficiently train the DCGANfor class-specific signal generation. In conclusion, the proposedhypergraph-based DA method has proved to be a competitive DAtechnique in comparison to other augmentation techniques
ViLU: Apprentissage d'incertitude langage et visuel pour la prédiction d'erreur
No full textInternational audienceReliable Uncertainty Quantification (UQ) and failure prediction remain open challenges for Vision-Language Models (VLMs). We introduce ViLU, a new Vision-Language Uncertainty quantification framework that contextualizes uncertainty estimates by leveraging all task-relevant textual representations. \ours constructs an uncertainty-aware multi-modal representation by integrating the visual embedding, the predicted textual embedding, and an image-conditioned textual representation via cross-attention. Unlike traditional UQ methods based on loss prediction, \ours trains an uncertainty predictor as a binary classifier to distinguish correct from incorrect predictions using a weighted binary cross-entropy loss, making it loss-agnostic. In particular, our proposed approach is well-suited for post-hoc settings, where only vision and text embeddings are available without direct access to the model itself. Extensive experiments on diverse datasets show the significant gains of our method compared to state-of-the-art failure prediction methods. We apply our method to standard classification datasets, such as ImageNet-1k, as well as large-scale image-caption datasets like CC12M and LAION-400M. Ablation studies highlight the critical role of our architecture and training in achieving effective uncertainty quantification.La quantification fiable de l'incertitude (UQ) et la prédiction des défaillances restent des défis à relever pour les modèles de vision-langage (VLM). Nous présentons ViLU, un nouveau cadre de quantification de l'incertitude visuelle-linguistique qui contextualise les estimations d'incertitude en exploitant toutes les représentations textuelles pertinentes pour la tâche. \ours construit une représentation multimodale sensible à l'incertitude en intégrant l'intégration visuelle, l'intégration textuelle prédite et une représentation textuelle conditionnée par l'image via une attention croisée. Contrairement aux méthodes UQ traditionnelles basées sur la prédiction des pertes, \ours forme un prédicteur d'incertitude en tant que classificateur binaire afin de distinguer les prédictions correctes des prédictions incorrectes à l'aide d'une perte d'entropie croisée binaire pondérée, ce qui le rend indépendant des pertes. Notre approche est particulièrement adaptée aux configurations post-hoc, où seuls les intégrations visuelles et textuelles sont disponibles sans accès direct au modèle lui-même. Des expériences approfondies sur divers ensembles de données montrent les gains significatifs de notre méthode par rapport aux méthodes de prédiction des défaillances de pointe. Nous appliquons notre méthode à des ensembles de données de classification standard, tels que ImageNet-1k, ainsi qu'à des ensembles de données d'images et de légendes à grande échelle, tels que CC12M et LAION-400M. Des études d'ablation soulignent le rôle essentiel de notre architecture et de notre formation dans la réalisation d'une quantification efficace de l'incertitude
A fully nonlinear cubic triangular multilayer Kirchhoff–Love shell element with accurate shear stress analysis
No full textA Correction to this paper has been published: https://doi.org/10.1007/s00466-025-02736-9International audienceThis paper presents a new triangular multi-layer nonlinear shell finite element, suitable for large displacements and rotations, and a formulation to obtain through the thickness shear stress considering both geometrical and material nonlinearity. This is a nonconforming element with 16 nodes, cubic displacement interpolation and enforcement of the rotation field based on Rodrigues rotation parameters and lagrangean parameters in 6 side nodes, with a total of 42 degrees of freedom. Associated with the new element, another novelty of this work is the development of a multilayer kinematical model with properties from Kirchhoff-Love theory, approximating the shell director across layers as constant and the formulation to obtain through the thickness shear stress considering both geometrical and material nonlinearity. The model is numerically implemented, and results are compared to different references in multiple examples, showing the capabilities of the formulation. Although lagrangean parameters are used to enforce C 1 continuity, it is believed that due to it being possibly the simplest multilayer extension, simple kinematic, a relatively small number of degrees of freedom, possibility to use various 3D material models, easily connected with multiple branched shells and beams, and geometric exact theory, this is a simple yet powerful shell element. The shear stress formulation, capable of being used with nonlinear materials, presents itself as a novelty, deviating from the standard use of linear material behaviour with co-rotational formulations, with good results
Proceedings of the 14th International Symposium on Imprecise Probabilities: Theories and Applications
No full textInternational audienc
Etudes environnementales, économiques et sociales de scénarios de gestion de biodéchets municipaux sur un territoire dans une logique d'économie circulaire
No full textInternational audienc
Robust Estimation of L1-Modal Regression Under Functional Single-Index Models for Practical Applications
No full textInternational audienceWe propose a robust procedure to estimate the conditional mode of a univariate outcome O given a Hilbertian explanatory variable I, under the assumption that (O,I) follow a single-index structure. The estimator is constructed using the M-estimator for the conditional density, and we establish its complete convergence. We discuss the estimator’s advantages in addressing challenges within functional data analysis, particularly robustness and reliability. We then evaluate both the performance and practical implementation of our method via Monte Carlo simulations. Furthermore, we carry out an empirical study to showcase the improved reliability and robustness of this estimator compared to conventional approaches. In particular, our methodology is applied to predict fuel quality based on spectrometry data, illustrating its strong potential in real-world scenarios
Extraction and chemical features of wood hemicelluloses: A review
No full textInternational audienceHemicelluloses have immense potential for applications in diverse fields, especially in polymeric materials. This review critically examines biomass treatment technologies, encompassing chemical, mechanical, and combined approaches to disrupt plant cell walls and enhance hemicellulose accessibility and solubility. The choice of a treatment method depends on factors like purpose, biomass composition, and economic and environmental considerations. Hemicelluloses extracted from wood are composed from up to 11 monomer units, most of them “neutral” monosaccharides (glucose, mannose etc.) but a couple “charged” (uronic acids). The average compositions of wood hemicelluloses change with the type of wood; the accuracy is not known. The content of “charged” monosaccharides may particularly suffer from underestimation due to strong hydrolysis. The chemical composition of intact wood hemicelluloses has never been determined: it is thus not known if hemicelluloses in wood are a mixture of several “simple” polysaccharides (such as glucomannans) or complex polysaccharides with up to 11 monomer units in the same macromolecule. Currently determined molecular weights (MW) of hemicelluloses range from 500 to 1,000,000 Da. The precision and accuracy of MWs are not known, primarily due to column inconsistency in analyzing anionic and neutral polymers. A combination of chromatography SEC methods and detectors is required for standardization