INRIA a CCSD electronic archive server
Not a member yet
122212 research outputs found
Sort by
Aligner méthode historique et RAG : transformer un assistant conversationnel en chaîne de preuve auditable et discutable
No full textCet article examine les défis suscitées par le déploiement de systèmes de Génération Augmentée par Récupération (RAG) dans l'exploration de sources historiques numérisées. Partant du constat d'une acceptabilité disciplinaire fragile, il pose la question suivante : comment garantir, avec un RAG appliqué à des archives bruitées et hétérogènes, des conditions de vérification et de critique compatibles avec la méthode historique ? Présenté comme un exposé de position, ce texte ne décrit pas un système stabilisé : il propose un cadrage et des pistes préliminaires pour orienter le développement de dispositifs RAG alignés avec ces exigences. Il propose de rétablir un contrôle sur la chaîne d'interprétation en articulant trois conditions : traçabilité (retrouver précisément documents et passages), auditabilité (rendre inspectables les transformations et paramètres de la chaîne), et discutabilité (mettre l'énoncé en débat en séparant preuve et interprétation). La contribution principale est une grille d'auditabilité traduisant des exigences historiennes en conditions instrumentées : (1) ancrage documentaire (provenance et intégrité), (2) séparation explicite citation/paraphrase/inférence, (3) restitution du contexte et pluralité des sources, (4) traçabilité des conditions d'exécution et diagnostic d'erreurs (récupération vs génération), ( 5) mécanismes d'abstention lorsque la preuve est insuffisante
HARCOM: Hardware complexity model for microarchitecture exploration
No full textInternational audienceMicroarchitecture exploration is generally conducted with performance simulators written in general-purpose programming languages, often C++. A performance simulator does not need to simulate all the details of the hardware implementation. It is often sufficient to simulate the events that can impact performance significantly, such as cache misses, branch mispredictions, data dependences, etc. Performance simulators often use approximations and abstractions. This is what allows them to simulate the execution of many instructions in a short amount of time, which is important for estimating millisecond-scale performance and for design space exploration.In general, microarchitects try to simulate realistic mechanisms. However, assessing the hardware complexity of a mechanism which only exists as a piece of C++ code in a performance simulator can be difficult. Hardware complexity is a multidimensional quantity including silicon area, energy consumption and delay. A simple, oft-used estimate of hardware complexity is the amount of storage used by a mechanism. Nevertheless, there is more to hardware complexity than storage. For instance, the delay of a branch predictor depends not only on its storage but also on the logic circuits processing the stored information. On the one hand, some hardware complexity models are available for microarchitecture research, such as CACTI and McPAT. However, their applicability is limited to cache-like structures or fixed microarchitectures. On the other hand, electronic design automation tools can be used to implement the hardware. However, this requires too much time and effort for microarchitecture exploration.HARCOM is a C++ library for estimating approximately the hardware complexity of microarchitectural parts, such as caches, branch predictors, hardware prefetcher, etc. HARCOM is compatible with existing performance simulators that are written in C++. HARCOM tries to find a useful middle ground between several contradictory objectives: the accuracy of the hardware complexity model, simulation speed, flexibility and ease of use. The microarchitectural part under study is modeled with HARCOM values instead of C++ integers. HARCOM simulates the functional behavior and, simultaneously, provides estimates of the silicon area, number of transistors, dissipated energy and circuits delays
A Quasi-Trefftz Method Based on Local Polynomial Impedance Boundary Conditions for Time-Harmonic Wave Propagation Problems
No full textFacing the need to solve time-harmonic wave problems in wide propagation domains, one needs to resort to iterative algorithms, to which classic numerical methods are usually poorly suited. The Trefftz method, based on the use of actual local solutions as basis functions, can then appear as a viable option for this kind of configurations. Thus, we first propose a general formalism allowing to gather different wave problems, so that a general Ultra Weak Variational Formulation can be defined for them at the same time. Yet, as its classic discretisation by plane waves is used to causing independence and accuracy limitations, we suggest to characterise local solutions by an impedance boundary conditions, which is discretised thanks to piecewise polynomial fields. Unfortunately, analytic expressions of such functions are not practically derivable, and we resort to a local numerical method so as to provide approximations: these quasi-solutions are finally chosen as basis functions in the Trefftz formulation, which is then referred as a quasi-Trefftz method. In the end, numerical experiments allow to highlight robust independence, convergence and iterative behaviour of this method, under calibration conditions of the local solver with respect to the polynomial BC discretisation
RooseBERT: A New Deal For Political Language Modelling
No full textThe increasing amount of political debates and politics-related discussions calls for the definition of novel computational methods to automatically analyse such content with the final goal of lightening up political deliberation to citizens. However, the specificity of the political language and the argumentative form of these debates (employing hidden communication strategies and leveraging implicit arguments) make this task very challenging, even for current general-purpose pre-trained Language Models. To address this issue, we introduce a novel pre-trained Language Model for political discourse language called Roose-BERT. Pre-training a language model on a specialised domain presents different technical and linguistic challenges, requiring extensive computational resources and large-scale data. RooseBERT has been trained on large political debate and speech corpora (8K debates, each composed of several sub-debates on different topics) in English. To evaluate its performances, we fine-tuned it on four downstream tasks related to political debate analysis, i.e., stance detection, sentiment analysis, argument component detection and classification, and argument relation prediction and classification. Our results demonstrate significant improvements over general-purpose Language Models on these four tasks, highlighting how domainspecific pre-training enhances performance in political debate analysis. We release Roose-BERT for the research community
Modeling and Parameter Estimation of Tumor Spheroids from Bottom-View Imaging: Beyond the Spherical Assumption
No full textTumor spheroids are widely used in vitro models for studying tumor growth and treatment response, yet their analysis is often limited by incomplete geometric measurements. In practice, only a two-dimensional bottom view is available, which can lead to strong biases when spheroids deviate from an ideal spherical shape, in particular for flattened aggregates. We propose a mechanistic modeling and data assimilation framework that accounts for both poor data quality and model uncertainty. A biologically motivated partial differential equations model with axial symmetry is introduced and reduced to a low-dimensional system compatible with sparse observations and ellipsoidal shapes. Uncertainties in initial conditions and parameters are handled using a Luenberger-type observer coupled with a reduced-order unscented Kalman filter. The approach is validated on synthetic data and applied to experimental in vitro spheroid data under propranolol treatment, illustrating how model-based integration of limited measurements improves the interpretation of spheroid growth dynamics and treatment-induced morphological changes
MRI contrast enhancement using schrodinger spectrum and unsupervised CNN: Amélioration du contraste IRM par le spectre de Schrödinger et un CNN non supervisé
No full textUltra-high-field (7T) MRI offers higher intrinsic SNR and resolution than lower-field MRI, yet images still suffer from spatially varying bias fields, intensity non-uniformity, and amplified high-frequency noise, causing local blur, poorer tissue separability, and potential diagnostic errors11. Prior work uses local contrast enhancement (e.g., CLAHE, histogram remapping) but these methods are sensitive to parameters and can over-enhance noise or create haloing/staircase artifacts; γ-based and SWT-BPJHE schemes can likewise co-amplify noise and ring edges at UHF2,3,4,52,3,4,5. We propose a post-acquisition enhancement based on 2D Semi-Classical Signal Analysis (SCSA), which represents images via eigenfunctions of a semi-classical Schrödinger operator and naturally suppresses noise6,76,7. To avoid manual, histogram-based selection, we learn a spatially varying, pixel-wise map G(x,y)G(x,y) with an unsupervised CNN88 and inject it into the SCSA remapping, yielding anatomically aware, adaptive contrast enhancement with reduced noise amplification
Gaussian Splatting Map Registration with Orthographic Bird's-Eye-View Renderings
No full textInternational audienceGaussian Splatting (GS) is a promising scene representation for visual localization and SLAM. Recent works have explored loop closure detection via Gaussian registration, improving map consistency and accuracy. However, achieving reliable registration given two GS representations from different acquisitions remains challenging. In this paper, we propose a complete pipeline to perform the matching and registration given two GS maps. The proposed method is grounded in generating orthographic bird's-eye views (BEVs) of optimized Gaussian models. The proposed approach leverages photometric and geometric information extracted directly from the GS to provide a trade-off of accuracy and invariance to different viewing changes (e.g., as types of GS maps, seasons, or illumination). Unlike existing 3D registration methods, which become inefficient as the number of Gaussians grows, our approach leverages 2D orthographic renders thus considerably reducing the registration complexity. Experiments on two public datasets demonstrate that our method achieves higher accuracy than several existing baselines, while also maintaining better registration results when dealing with GS maps learned by different techniques (e.g., 3DGS to LightGaussian), or GS maps presenting viewing changes such as varying illumination conditions.</div
Stratification of projection maps from toric varieties
No full textInternational audienceWe define a polyhedral version of a stratification for projection maps that applies to any complex or real toric variety and show that it yields similarly desirable properties to the classical map stratification of a proper map. Our results are constructive and give rise to a method for associating the Whitney strata of the projection to the faces of the polytope of the corresponding toric variety. For all the examples we consider, our resulting algorithm outperforms known general purpose methods, e.g., Helmer and Nanda (FoCM, 2022), and Đinh and Jelonek (DCG, 2021), for computing map stratifications
Novel genes arise from genomic deletions across the bacterial tree of life
No full textBacteria are hosts to enormous genic diversity. How new genes emerge, functionalize, and spread remain longstanding questions. Here, we explore a mechanism by which adaptive deletions fuse distant gene fragments. Unlike other gene birth mechanisms that begin with rare, neutral mutations, these "deletion-born fusions" reach high frequency by hitch-hiking on the deletion. The deletion-driven proliferation of the fusion prolongs the mutational supply within these genes before loss, providing additional opportunities for neofunctionalization. We document one such gene fixing and expressing in a long-term E. coli evolution experiment, and identify additional fusion events in the Mycobacterium tuberculosis-bovis split. Finally, we develop a scalable systematic screen to detect these genes in all 2.4 million public single-isolate genomes and identify deletion-born fusions across the bacterial tree of life. These findings challenge the notion that deletions are solely destructive and highlight their role as potential catalysts for evolutionary innovation
