Portail HAL des publications du LIRMM
Not a member yet
13279 research outputs found
Sort by
Overview of BirdCLEF+ 2025: Multi-Taxonomic Sound Identification in the Middle Magdalena, Colombia
No full textSource Agritrop Cirad (https://agritrop.cirad.fr/616719/)International audienceThe BirdCLEF+ 2025 challenge focused on the simultaneous acoustic identification of birds, amphibians, mammals and insects in the Middle Magdalena Valley, a biodiversity hotspot in Colombia. This edition aimed to advance passive acoustic monitoring by tasking participants with developing reliable systems for detecting and identifying multi-taxonomic vocalizations from extensive soundscape recordings. Using training data provided by museum collections, citizen science projects and new unlabeled soundscapes, participants addressed the challenge of out-of-distribution generalization under field conditions and limited training data for many species. Participants used data augmentation, pseudo-labeling, and self-training to enhance model robustness and accuracy, often refining pseudo-labels iteratively. For improved scores and runtime efficiency, teams commonly employed Test-Time Augmentation, ensemble methods, and optimized inference with dominant Sound Event Detection and CNN-based models, frequently pretraining on external datasets. The highest-scoring submission achieved an ROC-AUC score of 0.930 on the private leaderboard (0.933 on the public leaderboard), with the top 10 systems differing by only 0.9% in their scores
Beyond single drivers: A multi-stressor framework for understanding and managing jellyfish proliferation under concurrent anthropogenic and climate pressures
No full textInternational audienceThe Mediterranean Basin is a biodiversity hotspot increasingly threatened by anthropogenic pressures, such as coastalization, overfishing, and climate change. These stressors may contribute to jellyfish blooms, which jeopardize marine ecosystems and the services they provide. Yet, their combined effects remain poorly understood. This study models the responses of two jellyfish with contrasting thermal preferences, Rhizostoma pulmo and Aurelia spp., to global changes occurring in the Western Mediterranean Sea between 2004 and 2020. The loop analysis method was used to assess the combined effects of multiple anthropogenic pressures on the abundance of jellyfish and the model outputs were compared with the observed dynamics of both jellyfish. Results suggest that global change favored R. pulmo proliferation during the study period, whereas Aurelia spp. abundance did not show much change. The effects of different fishing effort scenarios on jellyfish blooms were also modeled to provide Western Mediterranean stakeholders with keys to remediation. While increased fishing effort in a context of global change is expected to foster jellyfish proliferation, reducing fishing pressure alone is unlikely to curb their abundance, because of the influence of concurrent antagonistic drivers. Our results highlight the multifactorial nature of jellyfish blooms and the need for integrated approaches to mitigate their proliferation. This study contributes to the objectives of the EU's Marine Strategy Framework Directive by improving the understanding of jellyfish as indicators of ecosystem imbalance and informing ecosystem-based management strategies
Déterminer un graphe à partir de son graphe de reconfiguration
No full textGiven a graph and a natural number , the -recolouring graph is the graph whose vertices are the -colourings of and whose edges link pairs of colourings which differ at exactly one vertex of . Recently, Hogan et al. proved that can be determined from provided is large enough (quadratic in the number of vertices of ). We improve this bound by showing that colours suffice, and provide examples of families of graphs for which colours do not suffice. We then extend this result to -Kempe-recolouring graphs, whose vertices are again the -colourings of a graph and whose edges link pairs of colourings which differ by swapping the two colours in a connected component induced by selecting those two colours. We show that colours suffice to determine in this case. Finally, we investigate the case of independent set reconfiguration, proving that in only a few trivial cases is one guaranteed to be able to determine a graph
Une approche pour optimiser le test de logiciel : deux exemples en algorithmique du texte
No full textMasterDans ce cours nous abordons la construction d'ensembles optimaux d'instances de tests pour des algorithmes centraux en algorithmique du texte et en bioinformatique
ScientiaRec: a Scientific Article Recommendation System with LLM-Driven Feature Extraction
No full textInternational audienceAccording to the European Commission, the number of scientific publications produced annually worldwide has more than tripled between 2000 and 2022. This rapid growth has made it increasingly difficult for researchers to identify and keep up with relevant work. Recommender systems (RSs) have emerged as a promising solution to this problem by helping users navigate the expanding scientific literature. In this paper, we introduce ScientiaRec, a novel scientific article RS that combines user-item interactions and content-based features, including descriptive tags and keywords automatically extracted using a large language model (LLM). At the core of our approach is a serendipity-aware matrix factorization model, designed to recommend relevant items while actively promoting serendipity. The goal is to help users discover novel and potentially insightful papers that go beyond their immediate research interests. We evaluate the performance of ScientiaRec against several baseline models using the publicly available CiteULike dataset, employing a comprehensive set of both accuracy-oriented and beyond-accuracy evaluation metrics. In addition, we conduct an LLM-based study to assess the serendipity of ScientiaRec Top-N recommendations. The experimental results demonstrate that ScientiaRec achieves a strong balance between relevance and beyond-accuracy objectives. Moreover, the inclusion of LLM-derived keywords significantly enhances the RS's overall performance
Analyzing Deep-Learning Methods for Power Line Component Detection in Unmanned Aircraft System Imagery With Few Data
No full textInternational audienceDue to the critical role of power lines in modern infrastructure, numerous automated methods have been developed for their inspection. Among these, Unmanned Aircraft Systems (UAS) have emerged as a valuable tool, offering rapid and precise inspections by capturing high-resolution aerial imagery of power lines. Drones enable access to hard-to-reach areas, reduce safety risks for workers near live wires, and significantly lower the time and cost associated with traditional inspection methods. In particular, deep learning techniques have been widely applied to automate the analysis of key components via the onboard camera. However, these methods typically rely on a first stage of detection based on large, annotated datasets focused on specific components, limiting their adaptability to new or unseen components. This paper investigates the application of two state-of-the-art algorithms of Few-Shot Object Detection (FSOD) for power line component detection: DeFRCN and CD-ViTO, alongside a modified Yolov8 detector of our own in which we integrated the modules of DeFRCN. We evaluate their performance using both public and proprietary datasets, analyzing unexpected outcomes and provide insights into the practical applicability of FSOD in real-world scenarios
A Scalable Two-Step Approach to Optimize Data and Energy Migrations in Mini Data Centers for Carbon-Neutral Computing
No full textInternational audienceThis poster presents an approach that combines fast VM migration algorithm with linear programming for energy allocation in computing systems. This approach significantly contributes to reducing computational time while supporting scalable and sustainable cloud operations
Developing a generic DSS model metadata catalogue and APIs for crop protection
No full textDecision Support Systems (DSS) in crop protection can give a helpful pest risk prognosis or recommendations for pest control, allowing farmers to make better-informed decisions. As a part of European Union policy strategy steering at the sustainable use of plant protection products, the EU H2020 project "IPM Decisions" created an online platform allowing farmers and advisors to access many DSS for major pests, weeds, and diseases in a variety of crops across Europe. The project has assessed and analyzed a set of research verified DSS´s for integrated pest management. Information on the models employed by these DSS has been collected into a model catalogue, which is used by the IPM Decisions platform.Two Application Programming Interfaces (APIs) have been built to facilitate use of these models. In the spirit of the FAIR (Findable, Accessible, Interoperable, Reproducible) principles, the DSS API provides access to DSS models with their metadata, including the description of model inputand output parameters. Weather API allows access to European on-line weather data sources and adapts their offering to the needs of the DSS models. The new APIs work as part of the IPM Decisions Platform, but they are publicly available so that also other agricultural software such as crop protection applications or farm management information systems (FMIS) can use them.In this article, we discuss the development of DSS and Weather APIs and present their structures and definitions. Finally, we present the services that DSS API and Weather API provide, and demonstrate the use of the APIs in three application cases
Growth Cost and Transport Efficiency Tradeoffs Define Root System Optimization Across Varying Developmental Stages and Environments in Arabidopsis
No full textCode availability: https://github.com/Salk-Harnessing-Plants-Initiative/AriadneInternational audienceRoot system architecture (RSA) is central to plant adaptation and fitness, yet the design principles and regulatory mechanisms connecting RSA to environmental adaptation are not well understood. We developed Ariadne, a semi-automated software for quantifying cost-efficiency tradeoffs of RSA by mapping root networks onto a Pareto-optimality framework, which describes the balance between resource transport efficiency and construction cost. Applying Ariadne to Arabidopsis thaliana , we found that root architectures consistently assume Pareto-optimal forms across developmental stages, genotypes, and environmental conditions. Using the Discovery Engine, an engine that combines machine learning together with interpretability techniques, we found developmental stage, the hy5/chl1-5 genotype, and manganese availability as important determinants of the cost-efficiency tradeoff, with manganese exerting a unique influence not observed for other nutrients. These results reveal that RSA plasticity is genetically constrained to cost-efficiency optimal configurations and that developmental and environmental factors shift RSA on the pareto front, with manganese acting as a strong modulator of the transport efficiency and construction cost balance
Statistical Quality and Reproducibility of Pseudorandom Number Generators in Machine Learning technologies
No full textMachine learning (ML) frameworks rely heavily on pseudorandom number generators (PRNGs) for tasks such as data shuffling, weight initialization, dropout, and optimization. Yet, the statistical quality and reproducibility of these generators-particularly when integrated into frameworks like PyTorch, TensorFlow, and NumPy-are underexplored. In this paper, we compare the statistical quality of PRNGs used in ML frameworks (Mersenne Twister, PCG, and Philox) against their original C implementations. Using the rigorous TestU01 BigCrush test suite, we evaluate 896 independent random streams for each generator. Our findings challenge claims of statistical robustness, revealing that even generators labeled "crush-resistant" (e.g., PCG, Philox) may fail certain statistical tests. Surprisingly, we can observe some differences in failure profiles between the native and framework-integrated versions of the same algorithm, highlighting some implementation differences that may exist