HAL-CentraleSupelec
Not a member yet
77624 research outputs found
Sort by
Le numérique frugal au service de la décarbonation
No full textParole d’experts : Groupe de Travail "Décarbonation & Numérique Frugal" de Systematichttps://systematic-paris-region.org/parole-dexperts-le-numerique-frugal-au-service-de-la-decarbonation-chapitre-5-lindispensable/Chapitre 5 : L’indispensable Stimuler et soutenir le développement des solutions logiciellesLien vers l'article (version web)Lien vers l'article (version pdf)</a
Geometry of Sparsity-Inducing Norms
No full textInternational audienceSparse optimization seeks an optimal solution with few nonzero entries. To achieve this, it is common to add to the criterion a penalty term proportional to the -norm, which is recognized as the archetype of sparsity-inducing norms. In this approach, the number of nonzero entries is not controlled a priori. By contrast, in this paper, our motivation is to find an optimal solution with at most~ nonzero coordinates (or for short, -sparse vectors), where is a given sparsity threshold (or ``sparsity budget''). For this purpose, we study the class of generalized -support dual~norms that arise from any given so-called source norm. When added as a penalty term, we provide conditions under which such generalized -support dual~norms promote -sparse solutions. The result follows from an analysis of the exposed faces of closed convex sets generated by -sparse vectors, and of how primal support identification can be deduced from dual information. Finally, we study some of the geometric properties of the unit balls for the -support dual~norms and their dual norms when the source norm belongs to the family of -norms. In particular, we show a striking structural property: every proper face of the unit balls for the -support dual~norms is a hypersimplex, i.e., the convex hull of -valued points with the same \lzero-norm
A stochastic use of the Kurdyka-Lojasiewicz property: Investigation of optimization algorithms behaviours in a non-convex differentiable framework
No full textInternational audienceAsymptotic analysis of generic stochastic algorithms often relies on descent conditions. In a convex setting, some technical shortcuts can be considered to establish asymptotic convergence guarantees of the associated scheme. However, in a non-convex setting, obtaining similar guarantees is usually more complicated, and relies on the use of the Kurdyka-Łojasiewicz (KŁ) property. While this tool has become popular in the field of deterministic optimization, it is much less widespread in the stochastic context and the few works making use of it are essentially based on trajectory-by-trajectory approaches. In this paper, we propose a new framework for using the KŁ property in a non-convex stochastic setting based on conditioning theory. We show that this framework allows for deeper asymptotic investigations on stochastic schemes verifying some generic descent conditions. We further show that our methodology can be used to prove convergence of generic stochastic gradient descent (SGD) schemes, and unifies conditions investigated in multiple articles of the literature
Bridging language models and knowledge graphs with controlled natural languages
No full textInternational audienceKnowledge graphs provide a source of up-to-date structured knowledge, which makes them an ideal counterpart to LLMs. LLMs, by themselves, are not trained to run structured queries internally and can become stale without a source of up-to-date information. We hypothesize that knowledge graphs can be effectively connected to large language models via controlled natural languages. Unlike standard formal query languages, controlled natural languages (CNLs) offer a syntax close to human language. Yet, can be unambiguously converted into formal languages such as SPARQL. In this article, we explore the premise that the extensive pre-training of LLMs on diverse textual data enables them to perform semantic parsing into controlled natural languages more accurately than parsing directly into formal query languages. To evaluate our hypothesis, we constructed a dataset facilitating the comparison between a standard formal language and two controlled natural languages. Our findings show a significant accuracy improvement when using the same amount of controlled natural language training samples. Additionally, fewer samples are required to achieve a desired performance when using CNLs compared to standard query languages. The higher data efficiency of CNLs is particularly important to reduce the complexity and cost of the collection and curation. This enables a more efficient way for LLMs to query KGs
Unmixing Algorithm in Multiplex Coherent Anti-Stokes Raman Scattering with Application to Label-Free Protein Discrimination
No full textCoherent Anti-Stokes Raman Scattering (CARS) microspectroscopy has emerged as a powerful technique for probing molecular assemblies with high chemical specificity. However, differentiating proteins within the same family remains a significant challenge. Here, we address the problem of the discrimination between actin and myosin filaments in human muscle. Starting with an inverse problem formulation, we propose a two-step computational resolution. First, the pure CARS spectra of actin and myosin are estimated. Second, these spectra are used to perform a spectral unmixing onto the whole image in order to quantify the relative contribution of each protein at the pixel level. Experiments on human muscle samples demonstrate that automated, label-free, and qualitative discrimination between protein families can be achieved using CARS microscopy
Performance of chitosan and alginate for soil stabilization under saturated conditions
No full textInternational audienc
Imagerie M-CARS exploitée hors résonances vibrationnelles pour la discrimination sans marquage de protéines musculaires
No full textInternational audienceLa discrimination sans marquage de familles de protéines dans une cible biomédicale constitue un défi majeur en imagerie optique non conventionnelle. Nous présentons une nouvelle approche d’optique non linéaire exploitant les données hyperspectrales multiplexées par diffusion Raman anti-Stokes cohérente (M-CARS), par excitation laser supercontinuum [1]. Les contributions résonnantes et non résonnantes du spectre CARS [2] sont exploitées pour discriminer sans marquage les réseaux d’actine et de myosine, principales familles de protéines du muscle squelettique [3]. Le signal non résonant, extrait de la région spectrale muette (entre 1700 et 2600 cm⁻¹) est généralement considéré comme non pourvoyeur d’information discriminante. La myosine est préalablement localisée par génération de seconde harmonique (SHG) délivrant ainsi les masques à partir desquels l’analyse pixel par pixel du rapport résonant/non résonant permet de discriminer la distribution de l’actine par rapport à la myosine avec une alternance et une périodicité attendue [4]. La Figure 1 illustre les différentes étapes du processus expérimental. Ces résultats montrent le potentiel de la région spectrale muetteet renforcent l’intérêt de l’imagerie CARS sans marquage pour les études biomédicales complexes
What is the carbon footprint of a 100% digital pathology scenario in France?
No full textInternational audienceHealthcare systems contribute to 5-10% of annual greenhouse gas (GHG) emissions and must reduce them to meet the 2016 Paris Agreement targets. Workflows in many areas of medicine are digitised, including surgical pathology. This is the case in France, the second most populous country in the European Union (68 million inhabitants/ 0•8% of world population) with an energy supply that is primarily sourced from nuclear power. We aimed to evaluate the carbon footprint (CF) in a theoretical scenario where all French surgical pathology laboratories implement full digitisation. Data on annual surgical pathology activities were obtained from the French National College of Pathologists. We defined the steps required by histological slide digitisation (scanning, image management software (IMS), data workflow, desktop tools, and storage) and inventoried the tools and their CF using a specific database and bibliography. This study was extrapolated to a national digitisation scenario without artificial intelligence (AI). Digitisation of a single French laboratory was also compared to the same laboratory, nondigitised. In 2021, 28,932,624 slides were generated, corresponding to more than 43PB (1•5 GB per whole-slide image (WSI)) shared by 250 surgical pathology laboratories using 500 scanners (i.e., two per laboratory working 12 hours per day). Digitising pathology resulted in 1,103 to 1,259 tons of carbon dioxide equivalent (CO2eq) for three months compared to 2,116 to 2,923 t CO2eq for one year of digital storage. This is equivalent to the GHG emissions of a combustion-powered car travelling around the Earth 145 times and 336 times, respectively. Data storage was the main contributor to the CF, regardless of storage type and duration (local or external for three months or one year). The other contributors included in decreasing order were desktop tools, scanners, and IMS. Finally, for one French surgical pathology laboratory, the CF of full digitisation increased by 4% to 8% (three months to one year of digital storage) compared to the current non-digitised CF. Such additional impact was equivalent to adding 13 to 26 minus 80-degree freezers in the laboratory. Full digitisation of surgical pathology has a significant environmental impact on global warming, even without the use of AI. Reducing storage duration and increasing the lifespan of digital equipment are the main strategies to mitigate these impacts
Assisting the early development stages of privacy-aware software: the PRIAM tooled metamodel for GDPR
No full textPart of a special issue on Regulatory compliance in software engineeringInternational audienceContext:As software systems are more tailored to users, personal data is collected and exploited more than ever before. This situation raises the issue of user privacy protection. Conforming to personal data protection regulations, such as the European General Data Protection Regulation (GDPR), has thus become a legal obligation for application providers. However, there are no widely adopted proposals to formalize, implement, and assess compliance with the personal data privacy protection required by GDPR.Objective:In order to help application developers in the early stages of the development process, our overarching objective is to propose a tooled software engineering approach to integrate personal data protection capabilities, thus contributing to the development-by-design of privacy-aware software aligned with GDPR requirements.Method:We developed a method called PRIAM (PRIvacy Assessment Method) that goes beyond a conceptual description of the regulation by incorporating concrete, actionable software artifacts. This article presents the cornerstone of this method – PRIAM metamodel – along with its companion artifacts.Results:PRIAM metamodel captures the main concepts of GDPR and is then supported by a domain-specific language, user stories, and a dedicated database schema. The comprehensiveness and relevance of PRIAM metamodel have been qualitatively evaluated by GDPR experts through a questionnaire. Complementarily, an AI-based evaluation has been conducted, using some Large Language Models (LLMs), opening perspectives for fast, iterative evaluations of metamodels that formalize regulation texts. Besides, the practicality and usefulness of PRIAM metamodel and all its companion artifacts are highlighted through the running example of a Sport center management application, where privacy enforcement features, tailored to the specific personal data of the application, are generated and integrated.Conclusion:These two elements assert the viability of our proposal as a practical solution for assisting the development of privacy-aware applications that are compliant with GDPR requirements, thanks to customizable sets of actual development artifacts, systematically derived from a validated comprehensive formalization of the regulation articles
Predicting rolling shear failure of cross-laminated panels using a multiscale domain decomposition approach
No full textInternational audienceCross-laminated timber (CLT) is gaining popularity as a sustainable building material due to its favorable structural performance and low environmental impact. This study introduces a multiscale domain decomposition method, known as the LaTIn strategy, to predict the out-of-plane behavior and failure of CLT panels. The proposed approach incorporates cohesive zone models to capture inter-and intralaminar damage using a cohesive parameter identification based on strength values reported in the literature. By aligning domain partitioning with principal stress directions and leveraging parallel computations, the method reduces degrees of freedom and the number of subdomains by up to 55% and 70%, respectively. Additional algorithmic enhancements allow for a 70% reduction of iterations. Validation with experimental data confirms the model's ability to predict rolling shear failure with high accuracy. These results demonstrate the potential of the strategy for efficient and reliable simulation of timber structures, in line with the needs of multiscale nonlinear analysis in structural engineering