Portail HAL Ensta
Not a member yet
11080 research outputs found
Sort by
Learning to Generate Training Datasets for Robust Semantic Segmentation
No full textInternational audienceSemantic segmentation techniques have shown significant progress in recent years, but their robustness to real-world perturbations and data samples not seen during training remains a challenge, particularly in safety-critical applications. In this paper, we propose a novel approach to improve the robustness of semantic segmentation techniques by leveraging the synergy between label-to-image generators and image-to-label segmentation models. Specifically, we design and train Robusta, a novel robust conditional generative adversarial network to generate realistic and plausible perturbed or outlier images that can be used to train reliable segmentation models. We conduct in-depth studies of the proposed generative model, assess the performance and robustness of the downstream segmentation network, and demonstrate that our approach can significantly enhance the robustness of semantic segmentation techniques in the face of real-world perturbations, distribution shifts, and out-of-distribution samples. Our results suggest that this approach could be valuable in safety-critical applications, where the reliability of semantic segmentation techniques is of utmost importance and comes with a limited computational budget in inference. We will release our code shortly
Understanding of water uptake mechanisms in an epoxy joint characterized by pore-type defects
No full textInternational audienceUnderstanding of water uptake mechanisms in an epoxy joint characterized by pore-type defects This work aims to characterize the water uptake mechanisms of a two-component epoxy adhesive joint immersed in deionized water. The pore-type defects in the bulk adhesive after the cure cycle are highlighted and characterized using X-ray µ-tomography. Two population patterns of defects are generated and analyzed, for two different thicknesses. The waterfront is not detectable by µ-tomography for this adhesive because the densities of the water and the adhesive remain relatively close to each other. Instead, the volume variation and kinetics of pore water filling have been accurately identified. This analysis was completed by optical observations and gravimetric measurements
Jack of All Trades, Master of Some, a Multi-Purpose Transformer Agent
No full textInternational audienceThe search for a general model that can operate seamlessly across multiple domains remains a key goal in machine learning research. The prevailing methodology in Reinforcement Learning (RL) typically limits models to a single task within a unimodal framework, a limitation that contrasts with the broader vision of a versatile, multi-domain model. In this paper, we present Jack of All Trades (JAT), a transformer-based model with a unique design optimized for handling sequential decision-making tasks and multi-modal data types. The JAT model demonstrates its robust capabilities and versatility by achieving strong performance on very different RL benchmarks, along with promising results on Computer Vision (CV) and Natural Language Processing (NLP) tasks, all using a single set of weights. The JAT model marks a significant step towards more general, cross-domain AI model design, and notably, it is the first model of its kind to be fully open-sourced at https://huggingface.co/jat-project/jat, including a pioneering general-purpose dataset
Dynamique ultrarapide dans les boites quantiques
No full textThis thesis focuses on the fundamental study and the optoelectronic applications of colloidal quantum dots of mercury calcogenide, with a preponderance for mercury telluride (HgTe). Thanks to the quantum confinement of the carriers, HgTe has incredible optical properties allowing it to modify its band gap and cover the entire infrared spectrum. In this thesis, we focused on a range between 2 and 4 microns, which is of great interest for guidance instruments but also for the detection of certain pollutants. This study was mainly carried out using femtosecond lasers, in the context of pump-probe spectroscopy experiments. This technique allowed the characterization of the electronic properties of the carriers that are essential to the development and realization of light sources and photodetectors. Indeed, the experimental data obtained in the framework of this thesis led to the realization of the first light-emitting LED centered at 2 microns as well as a photodetector centered at 4 microns and whose operating temperature could reach 200 K, unlike the 80 K of commercial mid-infrared photodetectors. Finally, this study also presents the behaviour of HgTe and monocrystalline bismuth as a function of temperature. The latter served as a standard sample for this part of the thesis.Cette thèse porte sur l’étude fondamentale et les applications optoélectroniques des boîtes quantiques colloïdales de calcogénide de mercure, avec une prépondérance pour le tellurure de mercure (HgTe). Celui-ci possède grâce au confinement quantique des porteurs, d’incroyables propriétés optiques lui permettant de modifier sa bande interdite et de couvrir l’intégralité du spectre infrarouge. Dans le cadre de cette thèse, nous nous sommes focalisés sur une gamme comprise entre 2 et 4 microns, qui présente un grand intérêt pour les instruments de guidage mais aussi la détection de certains polluants. Cette étude a principalement été réalisée par l’utilisation de lasers femtosecondes, dans le cadre de la réalisation d’expériences de spectroscopie pompe-sonde. Cette technique a permit la caractérisation des propriétés électroniques des porteurs qui sont indispensable au développement et à la réalisation de sources lumineuses ainsi que de photodétecteurs. En effet, les données expérimentales obtenues dans le cadre de cette thèse ont abouti à la réalisation de la première LED électroluminescente centrée à 2 microns ainsi que d’un photodétecteur centré à 4 microns et dont la température de fonctionnement a pu atteindre 200 K, à l’opposé des 80 K des photodétecteurs moyen infrarouge commerciaux. Pour finir, cette étude présente aussi le comportement du HgTe et du bismuth monocristallin en fonction de la température. Ce dernier nous ayant servi d’étalon pour cette partie de la thèse
Static and dynamic behaviour of ultra high molecular weight poly-ethylene (UHMWPE) Tensylon® composite
No full textInternational audienc
Modified error-in-constitutive-relation (MECR) framework for the characterization of linear viscoelastic solids
No full textInternational audienceWe develop an error-in-constitutive-relation (ECR) approach toward the full-field characterization of linear viscoelastic solids described within the framework of standard generalized materials. To this end, we formulate the viscoelastic behavior in terms of the (Helmholtz) free energy potential and a dissipation potential. Assuming the availability of full-field interior kinematic data, the constitutive mismatch between the kinematic quantities (strains and internal thermodynamic variables) and their ``stress'' counterparts (Cauchy stress tensor and that of thermodynamic tensions), commonly referred to as the ECR functional, is established with the aid of Legendre-Fenchel gap functionals linking the thermodynamic potentials to their energetic conjugates. We then proceed by introducing the modified ECR (MECR) functional as a linear combination between its ECR parent and the kinematic data misfit, computed for a trial set of constitutive parameters. The affiliated stationarity conditions then yield two coupled evolution problems, namely (i) the forward evolution problem for the (trial) displacement field driven by the constitutive mismatch, and (ii) the backward evolution problem for the adjoint field driven by the data mismatch. This allows us to establish compact expressions for the MECR functional and its gradient with respect to the viscoelastic constitutive parameters. For generality, the formulation is established assuming both time-domain (i.e. transient) and frequency-domain data. We illustrate the developments in a two-dimensional setting by pursuing the multi-frequency MECR reconstruction of (i) piecewise-homogeneous standard linear solid, and (b) smoothly-varying Jeffreys viscoelastic material
CLIP-QDA: An Explainable Concept Bottleneck Model
No full textInternational audienceIn this paper, we introduce an explainable algorithm designed from a multi-modal foundation model, that performs fast and explainable image classification. Drawing inspiration from CLIP-based Concept Bottleneck Models (CBMs), our method creates a latent space where each neuron is linked to a specific word. Observing that this latent space can be modeled with simple distributions, we use a Mixture of Gaussians (MoG) formalism to enhance the interpretability of this latent space. Then, we introduce CLIP-QDA, a classifier that only uses statistical values to infer labels from the concepts. In addition, this formalism allows for both local and global explanations. These explanations come from the inner design of our architecture, our work is part of a new family of greybox models, combining performances of opaque foundation models and the interpretability of transparent models. Our empirical findings show that in instances where the MoG assumption holds, CLIP-QDA achieves similar accuracy with state-of-the-art methods CBMs. Our explanations compete with existing XAI methods while being faster to compute
Fast, high-order numerical evaluation of volume potentials via polynomial density interpolation
No full textInternational audienceThis article presents a high-order accurate numerical method for the evaluation of singular volume integral operators, with attention focused on operators associated with the Poisson and Helmholtz equations in two dimensions. Following the ideas of the density interpolation method for boundary integral operators, the proposed methodology leverages Green's third identity and a local polynomial interpolant of the density function to recast the volume potential as a sum of single- and double-layer potentials and a volume integral with a regularized (bounded or smoother) integrand. The layer potentials can be accurately and efficiently evaluated everywhere in the plane by means of existing methods (e.g. the density interpolation method), while the regularized volume integral can be accurately evaluated by applying elementary quadrature rules. Compared to straightforwardly computing corrections for every singular and nearly-singular volume target, the method significantly reduces the amount of required specialized quadrature by pushing all singular and near-singular corrections to near-singular layer-potential evaluations at target points in a small neighborhood of the domain boundary. Error estimates for the regularization and quadrature approximations are provided. The method is compatible with well-established fast algorithms, being both efficient not only in the online phase but also to set-up. Numerical examples demonstrate the high-order accuracy and efficiency of the proposed methodology; applications to inhomogeneous scattering are presented
E-SCORE: A web-based tool for security requirements engineering
No full textInternational audienceAs digital systems continue to grow in popularity, they also become more vulnerable to various forms of attacks with various motives, including financial gain and political influence. In response, engineers must consider system security from the design phase. However, defining security requirements at this stage can be challenging. To address this challenge, we propose E-SCORE, a web-based tool that streamlines the security requirements engineering process. E-SCORE implements the SCORE (Security Criteria Ontology for security Requirements Engineering) (i) to suggest security mechanisms and additional criteria to enhance security coverage and (ii) to facilitate security analysis of the system. An example of banking system usage is provided. Through our approach, we could define ten additional security requirements for a single requirement. Therefore, E-SCORE offers a valuable resource for engineers to ensure the security of digital systems across various domains
