HAL Portal AU (University of Avignon)
Not a member yet
35713 research outputs found
Sort by
Late Fusion and Multi-Level Fission Amplify Cross-Modal Transfer in Text-Speech LMs
No full textText-Speech Language Models (TSLMs) -- language models trained to jointly process and generate text and speech -- are commonly trained through an early modality fusion/fission approach, in which both modalities are fed and predicted from a shared backbone via linear layers. We hypothesize that this approach limits cross-modal transfer by neglecting feature compositionality -- specifically, the finer-grained nature of speech representations compared to text -- preventing the emergence of a shared feature hierarchy within model layers. In this paper, we argue that this limitation can be addressed through late fusion and fission, with a fission process that accesses both high- and low-level features for speech generation. Our models implementing these principles, SmolTolk, rival or surpass state-of-the-art TSLMs trained with orders of magnitude more compute, and achieve significantly improved cross-modal performance relative to early fusion/fission baselines. Representation analyses further suggest that our method enhances the model's ability to abstract higher-level, more semantic features from speech, and leads to increasingly shared representation spaces across layers
Explaining machine learning metrics with bubbles and reverse correlation
No full textMachine learning et traitement du signal appliqués : acoustique et vibrations; GSAM - Acoustique Musicale: GVB - Vibro acoustique et Contrôle du Bruit: GAP - Voix et ParoleNational audienceMetrics are fundamental to evaluating models in machine learning, yet explaining the specific features driving high or low metric values often remains unexplored. This challenge is particularly evident in acoustics, where metrics evaluate models for soundscape classification, voice quality estimation, musical timbre perception simulation, and many other use cases in acoustics. Here, we present a method to address this issue, enabling "explainable metrics" by systematically identifying and visualizing the features responsible for variations in metric values. Initially inspired by psychophysical methods, this approach is based on the randomization of the input features of any machine learning model and the reverse analysis of model predictions. It provides deeper insights into how features influence key machine learning metrics such as r-squared, prediction accuracy, F1 score, or mean squared error, offering interpretability for complex models. While motivated by acoustic applications, this framework generalizes to diverse scientific domains, advancing transparency and understanding in metric-based analyses. In this presentation, I'll first introduce the method and present examples of applications in acoustics, such as for the characterization of sleepy voices, natural soundscapes, and musical instrument timbre
Partie IV : Cerner les modèles socioéconomiques dans le champ de l’insertionChapitre 1 : De l’entrepreneuriat social à un modèle d’affaires. Trajectoire d’une association d’insertion par le travail
No full textInternational audienc
14 Adapting Integrated Pest Management to Climate Change
No full textInternational audienceIntegrated pest management (IPM) is a holistic approach to pest control that combines various climate-dependent tactics. This chapter first reviews how climate, mainly temperature, influences four of these tactics: mating disruption, biological control using entomopathogens and entomophagous arthropods, and insecticides. Reports of recent failures in these techniques worldwide highlight the urgent need for adaptations to climate change, which are explored in the second section. The third section discusses recent and ongoing research efforts to ensure that IPM remains a viable production strategy in the context of climate change, using the codling moth Cydia pomonella (Lepidoptera:Tortricidae) as a case study
Innovative monitoring project to protect at risk archaeological heritage in Sudan: IMAHP Project: Innovative Monitoring Approaches for Heritage Protection
No full textInternational audienceActively involved in the protection of Sudan's archaeological sites, the SFDAS (French Archaeological Unit of Antiquities in Sudan) led the SAHPP project (Sudan Archaeological Heritage Protection Project) between 2020-2023 in collaboration with NCAM (National Corporation for Antiquities and Museums). Part of the project allowed the development of a GIS to centralize data on selected sites.Since the beginning of the war in 2023, the archaeological heritage has been severely exposed to destruction, looting and trafficking of cultural goods, making it difficult for the international scientific community to assess the extent of the damage.Faced with this situation, the SFDAS, in collaboration with the UMR 7300 ESPACE (Aix-Marseille University) has launched the IMAHP Project (Innovative Monitoring for Archaeological Heritage Protection in Sudan) based on the SAHPP data. It aims to develop a powerful tool for remote site monitoring, using Earth Observation System (EOS) coupled with an online GIS (WebSIG) and spatial analysis data for the identification, cartography, and modelling of risk exposures. A geospatial monitoring system combining artificial intelligence, archaeological, geographic, and multispectral remote sensing data will be developed to detect looting activities in near-real time. This system will enable the automated and predictive identification of the site's risk exposure, providing an innovative solution for the protection of Sudan's cultural heritage.The 27th Biennial Meeting of the Society of Africanist Archaeologists (SAfA 2025) will take place at the University of Algarve in Faro, Portugal. The event, hosted by ICArEHB – Interdisciplinary Centre for Archaeology and Evolution of Human Behaviour, will be held from the 21st to the 26th of July 2025.Link: https://safa2025.icarehb.com
Aux époque moderne et contemporaine. Un paysage essentiellement agricole
No full textInternational audienceChapitre du catalogue de l'exposition "Un passé incontournable. Découvertes archéologiques de l'A355", présentée au palais Rohan à Strasbourg (juin 2025-juin 2026
Variation of vegetation cover and the relationship with land surface temperature across Thailand (2007 to 2022)
No full textInternational audienceUnderstanding vegetation-climate interactions is essential amid escalating global climate change. This study investigates spatial-temporal and seasonal variations in Land Surface Temperature (LST) and Normalized Difference Vegetation Index (NDVI) across six regions of Thailand (2007-2022). Results reveal distinct regional and seasonal characteristics, with significant negative correlations between LST and NDVI (R = 0.61 dry; 0.39 rainy; 0.72 winter). The strongest negative correlation occurred during the rainy season in 2017, highlighting complex interannual variations. Seasonal LST fluctuations (winter-summer: 1.24, winter-rainy: -1.54, summer-rainy: -2.78, p < 0.001) and NDVI variations (winter-summer: 0.09, winter-rainy: 0.07, summer-rainy: -0.03, p < 0.001) were statistically significant. These findings emphasize monitoring LST and NDVI as vital for understanding ecological impacts of climate change and urbanization. The study specifically explores whether increased vegetation consistently is associated with lower temperatures, underscoring the importance of strategies to mitigate heat and enhance climate resilience, particularly in rapidly urbanizing regions
La mort en jeu. A propos de la mode du transi dans l’esthétique funéraire de la noblesse tardo-médiévale
No full textInternational audienc
Quid de l’environnement aux marges d’al-Andalus ? Approche interdisciplinaire des systèmes agrosylvopastoraux d’Albalat (Xe-XIIe s., Estrémadure, Espagne
No full textInternational audienceAfin de documenter l’histoire agricole des espaces marginaux et périphériques, une enquête paléoenvironnementale interdisciplinaire est actuellement menée sur le site archéologique d’Albalat (Xe-XIIe s., Estrémadure). La richesse des vestiges piégés sous les effondrements des habitations permet de restituer un large spectre des activités quotidiennes. La gestion des ressources, leur production et leur consommation sont abordées par des approches complémentaires : archéologique, bioarchéologique et chimique. Les résultats obtenus grâce à ces différentes approches permettent de documenter à la fois la production agricole (diversité des cultures, types de sols exploités, apports hydriques), la transformation des produits (battage des céréales, production de farine) et leur consommation (analyse des comblements de latrines, résidus organiques sur céramique). Ce travail propose une synthèse de ces résultats interdisciplinaires, dans l’objectif de qualifier ce système agrosylvopastoral situé aux marges d’al-Andalus