INRIA a CCSD electronic archive server
Not a member yet
122212 research outputs found
Sort by
Characterization of multifractality in oscillatory dynamics
No full textNeuronal oscillations exhibit large variability that has been hypothesized to reflect critical brain dynamics arising from the brains operating at a phase transition between disorder and order. A hallmark observable of such critical dynamics is scale-invariance that is typically quantified as long-range temporal correlations (LRTCs) in oscillation amplitude fluctuations. In this work, we extend the conventional scale-invariance-analysis framework for brain oscillations from selfsimilarity to multifractality. Using the new approach, we performed a multifractal analysis on the brain oscillation dynamics in different frequency bands of resting-state MEG recordings of 75 healthy human subjects. We estimated multifractal exponents from oscillation envelopes in frequencies from 5 to 96 Hz and assessed their anatomical and spectral patterns. To evaluate the statistical significance of multifractality estimates, we advance here two surrogate data generation methods. We found multifractality to be significant across the studied frequency range and to be most prevalent in the alpha-(8-14 Hz) and beta-(15-30 Hz) frequency bands. Moreover, we found self-similarity and multifractality estimates to be correlated and showed using the Principal Component Analysis that was explained by a correlation between the leading principal component and the multifractal estimates. Finally, we found the aggregate multifractal estimates to be highly consistent between sensor space and source-reconstructed MEG data
EcoWikiRS: Learning Ecological Representation of Satellite Images from Weak Supervision with Species Observations and Wikipedia
No full textInternational audienceThe presence of species provides key insights into the ecological properties of a location such as land cover, climatic conditions or even soil properties. We propose a method to predict such ecological properties directly from remote sensing (RS) images by aligning them with species habitat descriptions. We introduce the EcoWikiRS dataset, consisting of high-resolution aerial images, the corresponding geolocated species observations, and, for each species, the textual descriptions of their habitat from Wikipedia. EcoWikiRS offers a scalable way of supervision for RS vision language models (RS-VLMs) for ecology. This is a setting with weak and noisy supervision, where, for instance, some text may describe properties that are specific only to part of the species' niche or is irrelevant to a specific image. We tackle this by proposing WINCEL, a weighted version of the InfoNCE loss. We evaluate our model on the task of ecosystem zero-shot classification by following the habitat definitions from the European Nature Information System (EUNIS). Our results show that our approach helps in understanding RS images in a more ecologically meaningful manner
Multifractality in critical neural field dynamics
No full textThe brain criticality hypothesis has largely only characterized brain dynamics in terms of their self-similarity, although experimental evidence suggests that the brain exhibits significant multifractality. To understand how multifractality may emerge in critical-like systems modeling neuronal activity, we used a neural field model exhibiting neural oscillations and a critical phase transition. We find that multifractality emerges near a synchronization phase transition, and that the pattern of variation of multifractality changes when placing the model at a different phase transition. These findings show that multifractality in temporal dynamics emerges near criticality in neural fields, providing a formal basis for interpreting multifractality in brain recordings
Exploration Multi-Robot Distribuée avec Maintien de la Connectivité sous Contraintes de Qualité de Service
No full textMulti-Robot Systems (MRS) have become essential for autonomous missions in unknown or hazardous environments, notably in critical scenarios like search and rescue, where efficient mapping and robust communication are crucial. However, effectively balancing rapid exploration with reliable network connectivity remains challenging, especially for decentralized systems operating under dynamic conditions without centralized control. This thesis introduces a novel distributed multi-robot exploration algorithm, Dynamic Role-Based Exploration with Connectivity Maintenance (DRBECM), specifically designed to address these challenges. The proposed algorithm utilizes decentralized decision-making based on local information sharing, enabling autonomous role assignment among robots without relying on global information or centralized oversight. Robots dynamically adopt either "explorer" roles, focusing on maximizing information gain through frontier-based strategies, or "supporter" roles, employing flocking-inspired positioning to sustain robust communication links across the team. Neighbor selection and connectivity maintenance are efficiently managed using the Relative Neighborhood Graph (RNG). To further enhance exploration efficiency under realistic communication constraints, we extend DRBECM into a machine learning-enhanced framework, Multi-Robot Exploration via Flocking Coordination and Machine Learning-Driven Connectivity Assessment (DRBECM-ML). DRBECM-ML integrates distributed flocking dynamics with lightweight machine learning models, trained using real-world signal propagation data from the FIT-IoT-Lab testbed, to accurately predict Received Signal Strength Indicator (RSSI) values in real-time. These predictions significantly improve autonomous decision-making related to role-switching and frontier selection, ensuring stable and resilient communication networks throughout exploration tasks. Comparative evaluations indicate that tree-based algorithms, including Decision Trees and Extreme Gradient Boosting (XGBoost), offer optimal balance between prediction accuracy and computational efficiency suitable for deployment on mobile robots. Furthermore, this thesis investigates alternative communication strategies by comparing the performance of K-Nearest Neighbors (KNN) against the RNG for inter-robot communication, utilizing data generated via the Network Simulator 3 (NS-3) to analyze their effectiveness under various configurations. Our simulation results consistently show significant improvements in exploration time and reduction in redundant exploration compared to baseline approaches, while effectively maintaining network connectivity. Ultimately, this work aims to provide robust, adaptable, and decentralized multi-robot system solutions suitable for deployment in complex, dynamic, and infrastructure-limited real-world scenarios.Les Systèmes Multi-Robots (MRS) sont devenus essentiels pour les missions autonomes dans des environnements inconnus ou dangereux, notamment dans des scénarios critiques tels que la recherche et le sauvetage, où une cartographie efficace et une communication robuste sont cruciales. Cependant, équilibrer efficacement une exploration rapide avec une connectivité réseau fiable demeure un défi, particulièrement pour les systèmes décentralisés opérant dans des conditions dynamiques sans contrôle centralisé.Cette thèse présente un nouvel algorithme d'exploration multi-robots distribué, l'Exploration Dynamique Basée sur les Rôles avec Maintien de la Connectivité (DRBECM), spécifiquement conçu pour répondre à ces défis. L'algorithme proposé utilise une prise de décision décentralisée basée sur le partage d'informations locales, permettant une attribution autonome des rôles entre robots sans s'appuyer sur des informations globales ou une supervision centralisée. Les robots adoptent dynamiquement soit des rôles «d'explorateurs», se concentrant sur la maximisation du gain d'information à travers des stratégies basées sur les frontières, soit des rôles de «supporteurs», employant un positionnement inspiré du comportement d'essaim pour maintenir des liens de communication robustes à travers l'équipe. La sélection des voisins et le maintien de la connectivité sont gérés efficacement en utilisant le Graphe de Voisinage Relatif (RNG).Pour améliorer davantage l'efficacité d'exploration sous des contraintes de communication réalistes, nous étendons DRBECM en un framework amélioré par apprentissage automatique, DRBECM-ML (Exploration Multi-Robots via Coordination d'Essaim et Évaluation de Connectivité Pilotée par l'Apprentissage Automatique). DRBECM-ML intègre une dynamique d'essaim distribuée avec des modèles d'apprentissage automatique légers, entraînés en utilisant des données de propagation de signal du monde réel provenant de la plateforme d'essai FIT-IoT-Lab, pour estimer avec précision les valeurs d'Indicateur d'Intensité du Signal Reçu (RSSI) en temps réel. Ces prédictions améliorent significativement la prise de décision autonome liée au changement de rôles et à la sélection des frontières, assurant des réseaux de communication stables et résilients tout au long des tâches d'exploration. Les évaluations comparatives indiquent que les algorithmes basés sur les arbres, incluant les Arbres de Décision et l'Amplification de Gradient Extrême (XGBoost), offrent un équilibre optimal entre précision d'estimation et efficacité computationnelle adapté au déploiement sur robots mobiles.De plus, cette thèse examine des stratégies de communication alternatives en comparant les performances des k plus proches voisins (KNN) par rapport au RNG pour la communication inter-robots, utilisant des données générées via le simulateur réseau ns-3 pour analyser leur efficacité sous diverses configurations. Nos résultats de simulation montrent de manière constante des améliorations significatives du temps d'exploration et une réduction de l'exploration redondante comparé aux approches de référence, tout en maintenant efficacement la connectivité réseau. Enfin, ce travail vise à fournir des solutions de systèmes multi-robots robustes, adaptables et décentralisées, appropriées pour le déploiement dans des scénarios réels complexes, dynamiques et à infrastructure limitée
An operator approach to the analysis of electromagnetic wave propagation in dispersive media. Part 2: transmission problems.
No full textInternational audienceIn this second chapter, we analyse transmission problems between a dielectric and a dispersive negative material. In the first part, we consider a transmission problem between two half-spaces, filled respectively by the vacuum and a Drude material, and separated by a planar interface. In this setting, we answer the following question: does this medium satisfy a limiting amplitude principle? This principle defines the stationary regime as the large time asymptotic behavior of a system subject to a periodic excitation. In the second part, we consider the transmission problem of an infinite strip made of a Drude material embedded in the vacuum and analyse the existence and dispersive properties of guided waves. In both problems, our spectral analysis elucidates new and unusual physical phenomena for the considered transmission problems due to the presence of the dispersive negative material. In particular, we prove the existence of an interface resonance in the first part and the existence of slow light phenomena for guiding waves in the second part
Segmentation de nématodes dans des images de microscopie
No full textInternational audienceCe mémoire parle de la conception d’une application d’intelligence artificielle destinée à détecter et compter automatiquement les nématodes du pin à partir d’images de microscopie. J’y présente comment j’ai généré des données synthétiques, conçu un outil d’annotation pour les biologistes, puis entraîné et amélioré progressivement des modèles YOLO adaptés à cette tâche. L’application que j’ai développée permet de corriger facilement les prédictions, d’enregistrer les annotations et d’alimenter une boucle d’amélioration continue. Ce travail aboutit à un outil fonctionnel qui accélère le comptage en laboratoire et améliore la fiabilité des analyses
Quantum computing and artificial intelligence: status and perspectives
No full textThis white paper starts with describing how quantum computing could help develop innovative AI solutions, particularly in the ML spaces. This is a mid- to long-term effort. It is aligned with quantum computer hardware road maps. We then cover the use cases of classical AI to empower research and developments of quantum technologies, focused on quantum computing and quantum sensing. This application domain of AI will mature. One important aspect is to ensure classical AI scales well as the requirements of quantum computing platforms will grow, as the domain progressively shifts from NISQ devices to FTQCs. One exampleis the critical role of ML-powered quantum error correction (QEC) techniques [7]. At last, it provides a longer-term research agenda to drive work in foundational questions related to how AI and quantum computing interact and benefit each other. The white paper ends with a set of recommendations and challenges on the way to orchestrate the proposed theoretical work, align quantum AI developments with quantum hardware road maps, work on both classical and quantum resource estimates, particularly with the goal to mitigate and optimize energy consumption, orchestrate this upcoming hybrid software engineering discipline, and develop the European industry competitiveness while considering societal aspects
Fine-tuning des modèles Mistral 7B et 24B pour le domaine de la défense à l'aide d'un adaptateur QLoRA
No full textNational audienceThis document presents work on fine-tuning a Large Language Model (LLM) for the defense sector. The project relies exclusively on internal and external data provided by AMIAD (Agence Ministérielle pour l’IA de Défense - Ministerial Agency for Defense AI), including varied textual documents (PDF, Word, Wikipedia, etc.). The data preparation steps are detailed, as is the fine tuning strategy implemented on two models : Mistral 7B Instruct and Mistral-Small-24B-Instruct-2501. The report pays particular attention to the reproducibility of experiments and the management of carbon footprint.Ce document présente un travail de fine-tuning d'un modèle de langage (LLM) pour le secteur de la défense. Le projet repose exclusivement sur des données internes et externes fournies par l'AMIAD (Agence Ministérielle pour l'IA de Défense), incluant des documents textuels variés (PDF, Word, Wikipedia, etc.). Les étapes de préparation des données sont détaillées, de même que la stratégie de fine tuning mise en oeuvre sur deux modèles : Mistral 7B Instruct et Mistral-Small-24B-Instruct-2501. Le rapport accorde une attention particulière à la reproductibilité des expériences et à la gestion de l'empreinte carbone
Torque Observation of WRSM With Model Uncertainties for EV Applications
No full textInternational audienceIn this article, we propose a torque observation method based on a linear parameter varying (LPV) approach for a wound rotor synchronous machine (WRSM) used in electric vehicles (EVs), specifically for the Renault ZOE. The novelty of our approach lies in its ability to handle a wide range of uncertainties and parameter variations, such as speed fluctuations and model uncertainties in both magnetic flux and resistance. This enables more accurate and robust torque estimation, which is crucial for the demanding performance requirements of EV applications. We present a comprehensive observation methodology, which includes a state and unknown input observability study, robust LPV observer design, and a convergence analysis. The effectiveness of this approach is demonstrated through both simulations and experimental tests conducted on the BEMEVE real-power test bench. To highlight its merits, the performance of the LPV observer is compared to different types of observers
Herds From Video: Learning a Microscopic Herd Model From Macroscopic Motion Data
No full textInternational audienceFigure 1: Our method can simulate individual agents to replicate herd behaviour learnt from a video containing many animals. [Left and middle] The original video (lower-right) and our simulation (upper-left), optimized to fit macroscopic density and velocity fields over a coarse grid. [Right] An authored simulation in which a herd transitions between the two illustrated behaviours, featuring narrow and broad formations.</div