HAL-Paris 13
Not a member yet
33893 research outputs found
Sort by
Large Language Models Could Revolutionize Health Care, but Technical Hurdles May Limit Their Applications
No full textInternational audienc
Qsox1 Contributes to Vascular Remodelling in Response to Hypertension
No full textInternational audienceIntroduction: QSOX1, a sulfhydryl oxidase involved in arterial remodelling, has recently emerged as a biomarker for preeclampsia and acute heart failure. This study sought the cardiovascular roles of Qsox1 in response to angiotensin II (AngII)-induced hypertension. Methods: With approval from an Animal Ethics Committee (CNREEA#9), two models were developed: Qsox1-invalidated adult male mice (Qsox1−/−) mice (C57BL/6J background) and a tamoxifen-inducible, vascular smooth muscle cell (VSMC)-specific Qsox1 knockout. Hypertension was induced via AngII minipumps and trans-aortic constriction, with assessments of cardiac function, vessel size, and VSMC phenotype. Results: Qsox1−/− at baseline had lower blood pressure and exhibited a synthetic/immature VSMC phenotype in coronary arteries when compared to wild-type (WT). After 4 weeks of AngII infusion, Qsox1−/− mice showed acute heart failure, absent coronary media hypertrophy, and increased perivascular fibrosis compared to hypertensive WT controls (p < 0.01). VSMC-specific Qsox1 knockout leading to the lack of Qsox1 in VSMC only impairs the phenotype of these cells without effecting cardiac function in response to AngII. Conclusion: These data implicate vascular Qsox1 in the adaptive mechanisms of VSMC to pressure overload such as the development of media hypertrophy
Amélioration de l'efficacité des réseaux de neurones grâce aux connaissances a priori : normalisation, représentation multimodale et détection d'objets à vocabulaire ouvert
No full textThis research explores methods to improve deep neural networks by integrating priorknowledge, enhancing efficiency, and reducing reliance on large datasets and extensivecomputation. Key contributions include developing a framework to extend LLMs toMultimodal LLMs (MLLMs) at low cost, optimizing generative AI models for image recognition,NLP, and multimodal learning, and advancing domain adaptation and knowledge-drivengeneralization. The work also integrates efficient normalization techniques and modulararchitectures for scalable, cost-effective AI deployment.Cette recherche explore des méthodes pour améliorer les réseaux neuronaux profonds en intégrant des connaissances préalables, en augmentant l’efficacité et en réduisant la dépendance aux grands ensembles de données et aux calculs intensifs. Les principales contributions incluent le développement d’un cadre permettant d’étendre les LLMs aux LLMsmultimodaux (MLLMs) à faible coût, l’optimisation des modèles d’IA générative pour la reconnaissance d’images, le traitement du langage naturel (NLP) et l’apprentissage multimodal, ainsi que l’avancement de l’adaptation aux domaines et de la généralisation basée sur les connaissances. Ce travail intègre également des techniques de normalisation efficaces et des architectures modulaires pour un déploiement de l’IA évolutif et rentable
Extension of Intersection over Union to Improve Small Object Detection in Few-Shot Regime
No full textInternational audience—Intersection over Union (IoU) is a widely used criterion that quantifies the overlap between two bounding boxes. It plays a crucial role in object detection, serving both as a cost function for training detection models and as a criterion for evaluating their performance. The IoU value between the ground truth and predicted boxes determines whether a detection is considered accurate, based on a predefined threshold. However, this approach poses challenges when detecting small objects, especially in situations where annotated data is scarce, such as in few-shot learning scenarios. The scarcity of supervision hinders the learning of robust localization, which is especially detrimental for small objects. A small discrepancy of just a few pixels between the predicted and annotated bounding boxes can result in a false detection for small objects. To address these issues, we propose Scale-adaptive Intersection over Union (SIoU), a new controllable and adaptive similarity criterion that adjusts based on object size. First, SIoU helps to find a better balance between small and large objects during the training of few-shot detection methods, for which small objects are extremely problematic. Experiments on four distinct datasets show superior detection performance when using SIoU as a cost function. Second, by being more lenient with small objects, SIoU aligns more closely with human perception than IoU, making it a more suitable evaluation criterion
Détection de métaphores dans les documents médicaux
No full textNational audienceLa métaphore est une figure de style, qui permet de transférer le sens d'un terme source vers un terme cible, comme dans LE TEMPS C 'EST DE L 'ARGENT . De cette manière, la métaphore identifie des similarités cachées entre deux idées. La métaphore peut jouer plusieurs rôles dans la langue, comme l'embellir, structurer la pensée ou expliquer des notions complexes. Nous nous intéressons à la métaphore utilisée dans le domaine médical. Nous proposons d'abord une typologie de métaphores et un corpus de cas cliniques annoté avec des emplois métaphoriques. Nous effectuons également des expériences de détection automatique des métaphores avec un giga-modèle génératif. Plusieurs types de prompts sont testés. Les meilleurs résultats atteignent 67,50 de F-mesure, avec le rappel allant jusqu'à 74 % avec certains prompts . Le typage de métaphores montre que 45,51 % de métaphores sont typés correctement
SIMSAMU - Un jeu de données ouvert de dialogues de régulation médicale en français
No full textInternational audienceBackground: Dispatch Services (DS) are essential to Emergency Medical Services (EMS). Dispatchers enable patients to access medical assistance in emergencies, anytime and anywhere, within limited time and resources. AIbased decision-support tools hold great promise for dispatchers. Developing these tools requires medical fieldspecific data. Medical dispatch dialogue is unique: it is a brief phone exchange in an emergency, within a limited time frame, without a physical examination. Objective: Our main objective was to (i) create an open French dataset of medical dispatch dialogues. Our secondary objectives were to (ii) develop a detailed medical dispatch scheme from this dataset using an unsupervised method, and (iii) provide a baseline evaluation of diarization and speech recognition models for this domain in French. Methods: From 2022 to 2023, emergency medicine junior doctors simulated real-life medical dispatch calls. These calls were recorded and transcribed to form the SIMSAMU corpus. We developed a dispatch scheme based on (i) recording analysis, (ii) data-driven utterance typology, and (iii) domain expertise. Utterance typology was derived via hierarchical clustering of representations learned by finetuning BERT embeddings on SIMSAMU. Clusters were mapped to the Roter Interaction Analysis System (RIAS) and included in our dispatch scheme. SIMSAMU was used to train and evaluate state-of-the-art neural network models for diarization and speech recognition. Diarization used the PyaNet model, fine-tuned on the ESLO2 dataset. Speech recognition used a CTC model with pre-trained wav2vec 2.0 embedding, compared to the multilingual Whisper model. The CTC-wav2vec model was further fine-tuned on SIMSAMU and evaluated by leave-one-speaker-out cross-validation. Results: The dataset consists of 61 audio recordings totaling 3 h 14 min. Four clusters were identified for callers and 3 for dispatchers. Two main dialogue phases were identified: interrogation and contractualization. The diarization model achieved a 10.4 % error rate. Speech recognition word error rates were 35.8 % for Whisper, 24.8 % for the CTC-wav2vec model fine-tuned on ESLO2, and 16.1 % after in-domain fine-tuning. Conclusion:We propose a French open medical dispatch dialogue dataset and an expert-validated schema of the medical dispatch dialogue based on unsupervised analysis. Notable gaps in how well speech recognition models generalize underscore the need for targeted, in-domain fine-tuning in this specialized application. SIMSAMU is designed to support this effort by serving as a benchmark for evaluating domain-adapted speech recognition and dialogue modeling strategies.</div
L'exposition Slithering Cures - (Institut Français de Pondichéry, février 2024)
No full textDans un petit village du sud de l'Inde, chaque jour pendant la saison de la chasse, des dizaines de villageois adivasi appartenant à la communauté Irular (catégorisée dans les populations dites tribales) apportent des serpents parmi les plus dangereux du pays tels que cobras, vipères ou bongares, auprès d'un centre coopératif de collecte de venin, le seul d'Inde et géré par la communauté
Resurgence of fiscal interventionism: a longitudinal analysis of public aid to businesses in France since 1949
No full textThe 1980s marked the transition from demand-driven Fordist policies to supply-driven ones, emphasizing budgetary discipline and competitiveness. However, this shift did not signal the end of fiscal interventionism but rather its transformation. In this paper, we introduce a novel indicator that encompasses all government-to-business wealth transfers, including direct expenditures and tax expenditures. This new measure offers a clearer picture of the level of government support for businesses. Findings reveal that French public aid surged over three decades, reaching 8% of GDP by 2019, making it the fastest growing budget item since the 1990s. Fiscal policy has not been abandoned in the post-Fordist era; it is now employed as a supply-side strategy rather than a demand-side one. We also evaluate this fiscal policy, showing that while state aid has a limited impact on employment and investment, it significantly boosts corporate margins after the Great Financial Crisis
A stakeholder perspective of value creation versus value destruction under financialization: The cases of Bayer and Merck compared during 2000–2022
No full textInternational audienceOur political economy-based focus is on the value created within the non-financial corporation (NFC) under financialization, so that our interest is in productive value, not shareholder value (SHV). Indeed, we see both productive value generation and distributional conflict between stakeholders, as being particularly relevant to the case of Big Pharma. Here, arguably extraction and sometimes destruction of such value created (VC) are decided by the relative power of the various stakeholders. In this proposed methodology, we define and operationalize these concepts, tracking the development of such competing stakeholder claims on VC. We compare our resultant stakeholder-based and value-focused analysis of the Cash Flow statements of two German Big Pharma firms, Bayer and Merck, to others in the existing literature. Finally, we outline the theoretical and practical advantages of our proposal with its new metrics and argue that it provides an alternative highly necessary perspective contesting SHV-based approaches under financialization
Universal constructions in homotopical algebra
No full textWe apply the effective integration theory of Lie-graph algebras, developed recently by the authors, to the deformation and homotopy theories of types of bialgebras, that is structures controlled by a properad, like associative bialgebras, (involutive) Lie bialgebras, Frobenius bialgebras, double Poisson bialgebras, pre-Calabi--Yau algebras, quantum Airy structures, etc. In these cases, we provide their associated Deligne groupoid with an explicit homotopical description. We settle the Koszul hierarchy and the twisting procedure on the properadic level. We also give a conceptual construction of the homotopy transfer theorem in terms of gauge actions. This work extends the formulas for the deformation theory of operadic algebras