INRIA a CCSD electronic archive server
Not a member yet
122212 research outputs found
Sort by
EuroBERT: Scaling Multilingual Encoders for European Languages
No full textInternational audienceEqual contribution, † Ordered alphabetically by the first name, ⋆ Senior advisor General-purpose multilingual vector representations, used in retrieval, regression, and classification, are traditionally obtained from bidirectional encoder models. Despite their wide applicability, encoders have been recently overshadowed by advances in generative decoder-only models. However, many innovations driving this progress are not inherently tied to decoders. In this paper, we revisit the development of multilingual encoders through the lens of these advances, and introduce EuroBERT, a family of multilingual encoders covering European and widely spoken global languages. Our models outperform existing alternatives across a diverse range of tasks, spanning multilingual capabilities, mathematics, and coding, and natively support sequences of up to 8,192 tokens. We also examine the design decisions behind EuroBERT, offering insights into our dataset composition and training pipeline. We publicly release the EuroBERT models, including intermediate training checkpoints, together with our training framework
The medial axis of any closed bounded set is locally Lipschitz stable with respect to the Hausdorff distance under ambient diffeomorphisms
No full textInternational audienceWe prove that the medial axis of a closed set is locally Hausdorff stable in the following sense: Let S ⊆ R d be a fixed compact set and S(c, r ) some sphere with radius r containing S in its interior. Consider the space of C 1,1 diffeomorphisms of R d to itself, which keep the exterior of S(c, r ) invariant. The map from this space of diffeomorphisms (endowed with a Banach norm) to the space of closed subsets of R d (endowed with the Hausdorff distance), mapping a diffeomorphism F to the closure of the medial axis of F(S ∪ S(c, r )), is Lipschitz. A similar statement holds if S is non-compact but the Hausdorff distance between S and R d is bounded; in other words, every point in R d has a point in S nearby. The latter statement can further be localized at the cost of having to consider two one-sided Hausdorff distances. Our result extends result of Chazal and Soufflet on the stability of the medial axis of C 2 manifolds under C 2 ambient diffeomorphisms
Burning or Collapsing the Medial Axis is Unstable
No full textInternational audienceThe medial axis of a set consists of the points in the ambient space without a unique closest point in the original set. Since its introduction, the medial axis has been used extensively in many applications as a method of computing a skeleton topologically equivalent to the original set. Unfortunately, one limiting factor in the use of the medial axis of a smooth manifold is that it is not necessarily topologically stable under small perturbations of the manifold. To counter these instabilities, various prunings of the medial axis have been proposed in the computational geometry community. Here, we examine one type of pruning, called burning. Because of the good experimental results it was hoped that the burning method of simplifying the medial axis would be stable. In this work, we show a simple example that dashes such hopes. Based on Bing's house with two rooms, we demonstrate an isotopy of a shape where the medial axis goes from collapsible to non-collapsible. More precisely, we consider the standard deformation retract from the closed ball to Bing's house with two rooms, but stop just short of the point where Bing's house becomes two dimensional. This way we obtain an isotopy from the 3-ball to a thickened version of Bing's house. Under this isotopy, the medial axis goes from collapsible to non-collapsible. We stress that this isotopy can be made generic, in the sense of singularity theory, as developed by Arnol'd and Thom
AMStraMGRAM: Adaptive Multi-cutoff Strategy Modification for ANaGRAM
No full textRecent works have shown that natural gradient methods can significantly outperform standard optimizers when training physics-informed neural networks (PINNs). In this paper, we analyze the training dynamics of PINNs optimized with ANaGRAM, a natural-gradient-inspired approach employing singular value decomposition with cutoff regularization. Building on this analysis, we propose a multi-cutoff adaptation strategy that further enhances ANaGRAM's performance. Experiments on benchmark PDEs validate the effectiveness of our method, which allows to reach machine precision on some experiments. To provide theoretical grounding, we develop a framework based on spectral theory that explains the necessity of regularization and extend previous shown connections with Green's functions theory
Conformal Prediction for Long-Tailed Classification
No full textMany real-world classification problems, such as plant identification, have extremely long-tailed class distributions. In order for prediction sets to be useful in such settings, they should (i) provide good class-conditional coverage, ensuring that rare classes are not systematically omitted from the prediction sets, and (ii) be a reasonable size, allowing users to easily verify candidate labels. Unfortunately, existing conformal prediction methods, when applied to the long-tailed setting, force practitioners to make a binary choice between small sets with poor class-conditional coverage or sets with very good class-conditional coverage but that are extremely large. We propose methods with guaranteed marginal coverage that smoothly trade off between set size and class-conditional coverage. First, we propose a conformal score function, prevalence-adjusted softmax, that targets a relaxed notion of class-conditional coverage called macro-coverage. Second, we propose a label-weighted conformal prediction method that allows us to interpolate between marginal and class-conditional conformal prediction. We demonstrate our methods on Pl@ntNet and iNaturalist, two long-tailed image datasets with 1,081 and 8,142 classes, respectively
QUARTZ : Approche abstractive non supervisée par question-réponse pour le résumé de dialogue orienté tâche
No full textInternational audienceDialogue summarization condenses conversations into concise text, reducing dialogue complexity in dialogue-heavy applications. Existing approaches heavily rely on costly human-written data, and the resulting summaries often lack task-specific focus, leading to suboptimal performancefor downstream tasks, such as medical ones. In this paper, we introduce QUARTZ, a framework for task-oriented unsupervised dialogue summarization. QUARTZ starts by generating multiple summaries and task-specific question-answer pairs using large language models (LLMs). Summariesare evaluated by having the LLMs respond to task-related questions before (i) selecting the best candidate responses and (ii) identifying the most informative summary. Finally, we finetune the best LLM on the selected summaries. When validated on multiple datasets, QUARTZ achieves competitivezero-shot performance, rivaling fully-supervised State-of-the-Art (SoTA) approaches.Le résumé de dialogues condense les conversations en un texte concis, réduisant la complexité des applications riches en interactions. Les approches existantes reposent souvent sur l’entraînement de modèles de langue à imiter des résumés humains. Cependant, cette approche est coûteuse etles résumés obtenus manquent souvent de pertinence, entraînant des performances sous-optimales, notamment en médecine. Dans cet article, nous introduisons QUARTZ, une méthode non supervisée pour le résumé de dialogues orienté tâche. QUARTZ génère plusieurs résumés et paires de questionsréponses à l’aide de grands modèles de langue (LLMs). Les résumés sont évalués en demandant aux LLMs de répondre à ces questions avant (i) de sélectionner les meilleures réponses et (ii) d’identifier le résumé le plus informatif. Enfin, nous affinons le meilleur LLM sur les résumés générés sélectionnés. Validé sur plusieurs ensembles de données, QUARTZ atteint des performances compétitives en zéro-shot, rivalisant avec les approches supervisées de pointe
Overlapping data in network protocols: bridging OS and NIDS reassembly gap
No full textInternational audienceIPv4, IPv6, and TCP have a common mechanism allowing one to split an original data packet into several chunks. Such chunked packets may have overlapping data portions and, OS network stack implementations may reassemble these overlaps differently. A Network Intrusion Detection System (NIDS) that tries to reassemble a given flow data has to use the same reassembly policy as the monitored host OS; otherwise, the NIDS or the host may be subject to attack. In this paper, we provide several contributions that enable us to analyze NIDS resistance to overlapping data chunks-based attacks. First, we extend state-of-the-art insertion and evasion attack characterizations to address their limitations in an overlap-based context. Second, we propose a new way to model overlap types using Allen's interval algebra, a spatio-temporal reasoning. This new modeling allows us to formalize overlap test cases, which ensures exhaustiveness in overlap coverage and eases the reasoning about and use of reassembly policies. Third, we analyze the reassembly behavior of several OSes and NIDSes when processing the modeled overlap test cases. We show that 1) OS reassembly policies evolve over time and 2) all the tested NIDSes are (still) vulnerable to overlap-based evasion and insertion attacks
De l'ascendance bioculturelle du design d'interaction
No full textInternational audienceNotre hypothèse épistémologique d'une ascendance bioculturelle du design d'interaction vise à inscrire la discipline dans la profonde généalogie des activités de l'espèce humaine. Notre intention découle de la nécessité d'une clarification des origines et du sens de cette jeune discipline, dont les savoirs, les savoir-faire, l'enseignement et l'exercice manquent de structure et de cohérence. Afin d'étayer notre hypothèse, nous passons en revue les déterminants pluri-disciplinaires qui fondent la discipline et régissent ses produits, et nous nous attardons sur les qualités du mapping des données d'entrée issues d'interfaces tangibles dont l'expressivité atteint un degré organique. Notre proposition théorique dégage des éléments de déontologie, de cadrage des conditions d'existence et de validité des savoirs et savoir-faire de la discipline à destination de ses praticiens, enseignants et apprenants
Le rôle de l'interaction sociale dans la découverte des raccourcis clavier
No full textNational audienceNous étudions comment la méthode des raccourcis clavier est découverte, une condition nécessaire pour leur adoption. Nous avons réalisé une enquête rétrospective à grande échelle (N=853) visant à recueillir les récits de la première utilisation des raccourcis clavier (1) Les personnes interrogées déclarent souvent qu'une interaction sociale est à l'origine de la découverte, et peu d'entre elles déclarent avoir utilisé des aides informatiques telles que des infobulles ; (2) les personnes interrogées découvrent les raccourcis clavier à l'école primaire ou au collège (entre 6 et 14 ans), bien que cela dépende de la politique éducative locale ; (3) la découverte se fait généralement avec un enseignant, un camarade de classe ou un membre de la famille pendant les devoirs à la maison ; et (4) les personnes interrogées font souvent état d'une forte réaction émotionnelle à la suite de cette découverte
Unsupervised Parallel Physics Informed Neural Networks for Blood Flow Estimation and Cardiovascular Parameters Assessment ⋆
No full textThis study presents a comprehensive investigation into the estimation of blood flow and cardiovascular parameters using physics-informed neural networks (PINNs) coupled with ensemble learning techniques. Four distinct architectures were investigated and implemented, incorporating various Windkessel models. The Windkessel model is a cardiovascular model linking blood flow to blood pressure and is categorized as a lumped (0D) model. We explored various Windkessel models: the two-and three-element Windkessels, as well as their fractional versions, which include fractionalorder capacitors. A complete physics-informed deep-learning estimation framework is proposed, taking as input blood pressure signals and giving as output estimated blood flow and cardiovascular parameters. Through rigorous training and fine-tuning, robust convergence was achieved, enabling accurate estimation of blood flow and cardiovascular parameters. Using two datasets for validation with various data distributions and patients' clinical conditions, the results demonstrate consistent performance across various cardiac profiles. The transferability of the trained models for unseen data distribution was investigated. Particularly, the introduction of fractional order elements in the Windkessel models exhibited improved accuracy without increasing model complexity. The analysis of parameters' distributions and their physiological coherence further validated the reliability of the obtained estimations. Precisely, we conducted an in-depth study of the estimated peripheral vascular resistance and fractional order, comparing them to their expected values according to the literature. A comparative study between the proposed method and existing literature showed the efficiency of the proposed framework. Finally, the framework is validated using In-Vivo human data to demonstrate its effectiveness on real-world measurements.</div