Portail HAL des publications du LIRMM
Not a member yet
13279 research outputs found
Sort by
Spectral Conditions for the Ingleton Inequality
No full textThe Ingleton inequality is a classical linear information inequality that holds for representable matroids but fails to be universally valid for entropic vectors. Understanding the extent to which this inequality can be violated has been a longstanding problem in information theory. In this paper, we show that for a broad class of jointly distributed random variables (X,Y) the Ingleton inequality holds up to a small additive error, even even though the mutual information between X and Y is far from being extractable. Contrary to common intuition, strongly non-extractable mutual information does not lead to large violations of the Ingleton inequality in this setting. More precisely, we consider pairs (X,Y) that are uniformly distributed on their joint support and whose associated biregular bipartite graph is an expander. For all auxiliary random variables A and B jointly distributed with (X,Y), we establish a lower bound on the Ingleton quantity I(X:Y|A)+I(X:Y|B)+I(A:B) - I(X:Y) in terms of the spectral parameters of the underlying graph. Our proof combines the expander mixing lemma with a partitioning technique for finite sets
Design and Optimization of Solar-Powered Embedded Systems with Uppaal Stratego
No full textInternational audienceEnergy intermittency in solar-powered embedded systems threatens Quality of Service (QoS) and system autonomy. In this study, we address the design of these systems with a formal co-design approach that provides verifiable guarantees, a critical advantage over traditional heuristic or predictive methods that often fail under unpredictable conditions. We use timed automatabased modeling within Uppaal Stratego to minimize grid reliance and battery capacity in a typical system, under QoS guarantees. Our methodology demonstrates that synthesized control strategies can reduce grid reliance by 58-72%, while an optimized task scheduling heuristic can decrease required battery capacity by up to 13% compared to the baseline. Our approach provides a formal basis for comparing these techniques to inform system design
Stable Vision-Based Robot Kinematic Control with Deep Learning-Based Oriented Object Detector
No full textInternational audienceRecent advances in machine learning and deep learning have significantly enhanced robot control by improving object detection and visual feature extraction. However, ensuring theoretical guarantees of stability and convergence in learning-enabled control systems remains a major challenge. In this paper, we propose a vision-based control framework that integrates a deep learning oriented-object detector with a Lyapunov-stable servo control law. The proposed method ensures provably stable convergence of the robot end-effector or its grasped object's pose to a desired camera image region for both eye-in-hand and eye-to-hand configurations. Unlike existing deep learning based visual servoing methods, which either lack formal stability guarantees or ignore object orientation control, our approach incorporates object orientation into the control loop through a region-based method using quaternion representation and formally guarantees stability. We validated our framework on a 6-DoF UR5e manipulator performing cup insertion and centering tasks, demonstrating accurate and stable control in both camera setups
Spatiotemporal dynamics of computed tomography perfusion (CTP) parameters following aneurysmal subarachnoid hemorrhage (SAH)
No full textInternational audiencePreclinical and clinical studies suggest an early and progressive deterioration in cerebral perfusion leading to delayed cerebral ischemia (DCI) after aneurysmal subarachnoid hemorrhage (SAH). Computed Tomography Perfusion (CTP) has then been investigated for DCI diagnosis and prognosis. However, spatial and temporal evolution of CTP-derived metrics have not been established, such that optimal CTP periodicity for monitoring and metrics thresholds for triggering intervention remain unclear. We developed an image processing pipeline to quantify CTP parameters' dynamics in SAH patients. Sixty-two patients were retrospectively included. Cerebral vasospasm (CVS) occurrence was 68%, that of DCI 15%. CTP parameters displayed specific dynamics in each of the noCVS, CVS, and DCI groups. In the DCI category, CBF showed early hyperemia and global flow homogenization, followed by an increased mean and variability of TMax. These features were included in a DCI predictive model (AUC = 0.94 after bootstrapping correction). Two types of dynamics emerged in DCI patients, one characterized by high asymmetry between hemispheric parameters, the other by a rapid whole-brain deterioration of brain perfusion. In conclusion, CTP parameters' early dynamics allow to sort SAH patients that will develop either CVS or DCI, advocating for repeating CTP examinations to adapt therapeutic strategies
Cooperative learning of Pl@ntNet's Artificial Intelligence algorithm: how does it work and how can we improve it?
No full textInternational audienceDeep learning models for plant species identification rely on large annotated datasets. The PlantNet system enables global data collection by allowing users to upload and annotate plant observations, leading to noisy labels due to diverse user skills. Achieving consensus is crucial for training, but the vast scale of collected data makes traditional label aggregation strategies challenging. Existing methods either retain all observations, resulting in noisy training data or selectively keep those with sufficient votes, discarding valuable information. Additionally, as many species are rarely observed, user expertise can not be evaluated as an inter-user agreement: otherwise, botanical experts would have a lower weight in the AI training step than the average user. Our proposed label aggregation strategy aims to cooperatively train plant identification AI models. This strategy estimates user expertise as a trust score per user based on their ability to identify plant species from crowdsourced data. The trust score is recursively estimated from correctly identified species given the current estimated labels. This interpretable score exploits botanical experts' knowledge and the heterogeneity of users. Subsequently, our strategy removes unreliable observations but retains those with limited trusted annotations, unlike other approaches. We evaluate PlantNet's strategy on a released large subset of the PlantNet database focused on European flora, comprising over 6M observations and 800K users. We demonstrate that estimating users' skills based on the diversity of their expertise enhances labeling performance. Our findings emphasize the synergy of human annotation and data filtering in improving AI performance for a refined dataset. We explore incorporating AI-based votes alongside human input. This can further enhance human-AI interactions to detect unreliable observations
Robustness of high‐throughput prediction of leaf ecophysiological traits using near infrared spectroscopy and poro‐fluorometry
No full textData availability statement: The datasets supporting the conclusions of this article are available in the Recherche Data Gouv repository at https://doi.org/10.57745/WVAPOL. The scripts for prediction analyses are publicly available in GitLab: https://forgemia.inra.fr/eva.coindre/robustness_ht_prediction_ecophysiological_traits.International audienceAbstract:Water scarcity is a major threat to crop production and quality. Improving drought tolerance through variety selection requires a deeper understanding of plant ecophysiological responses, but large-scale phenotyping remains a bottleneck. This study assessed the potential of high-throughput tools (spectroscopy and poro-fluorometry) to predict leaf morphological and ecophysiological traits in a grapevine diversity panel grown in pots under well-watered outdoor conditions and under three contrasting soil water treatments in a greenhouse. We found a certain complementarity between measuring devices. Spectrometers could accurately predict leaf mass per area, water content, and water quantity (R2 > 0.58), while the poro-fluorometer was efficient for predicting net CO2 assimilation (R2 > 0.72), regardless of the water treatment. The prediction of leaf mass per area using spectrometers appeared to be quite robust across both outdoor and greenhouse experiments, while the prediction of water use efficiency was dependent on the water treatment, with much better predictions under moderate (R2 > 0.73) than severe water deficit. Calibrated models were then applied to the full diversity panel using only high-throughput measurements to estimate trait values and their broad-sense heritability. Leaf mass per area, also measured directly, showed similar heritability whether based on observed or predicted data. Heritability estimates for predicted traits reached up to 0.5. Overall, our findings support the use of spectroscopy and poro-fluorometry as reliable, nondestructive tools for high-throughput phenotyping, enabling genetic studies on drought-related traits in grapevine.Plain Language Summary: Drought is a major challenge for crop production. To breed grapevines that can better tolerate dry conditions, it is crucial to evaluate how plants respond to water stress using quick phenotyping tools. In this study, we tested two fast, nondestructive tools, near-infrared spectroscopy and poro-fluorometry, on grapevines grown with different water levels. Spectroscopy was accurate for evaluating leaf thickness and water content, while poro-fluorometry was better at predicting photosynthesis. These tools were effective even across different growing environments. The results demonstrate that these methods can aid in estimating genetic variability in drought-related traits, facilitating the selection and improvement of drought-tolerant grapevines
The robust selection problem with information discovery
No full textInternational audienceWe explore a multiple-stage variant of the min-max robust selection problem with budgeted uncertainty that includes queries. First, one queries a subset of items and gets the exact values of their uncertain parameters. Given this information, one can then choose the set of items to be selected, still facing uncertainty on the unobserved parameters. In this paper, we study two specific variants of this problem. The first variant considers objective uncertainty and focuses on selecting a single item. The second variant considers constraint uncertainty instead, which means that some selected items may fail. We show that both problems are NP-hard in general. We also propose polynomial-time algorithms for special cases where the sets of items that can be simultaneously queried are defined by a cardinality or a knapsack constraint. For the problem with constraint uncertainty, we also show how the objective function can be expressed as a linear program, leading to a mixed-integer linear programming reformulation for the general case. We illustrate the performance of this formulation using numerical experiments
Conformal Prediction for Long-Tailed Classification
No full textInternational audienceMany 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
Unsupervised data-driven detection of exceptional atmospheric trajectories
No full textExtreme weather events in Europe are closely linked to the large-scale atmospheric circulation and often develop over several consecutive days. Most existing circulation-based approaches focus on identify extreme weather patterns as instantaneous atmospheric states and therefore do not explicitly account for the temporal evolution of the flow. In this study, we apply a data-driven and unsupervised methodology to identify rare atmospheric trajectories from reanalysis data. The method quantifies how isolated short segments of atmospheric evolution are within the space of all observed trajectories, using daily sea-level pressure fields over Europe. We apply the approach to several decades of reanalysis data and identify the most isolated trajectories for different trajectory lengths. The detected trajectories are characterised by large-scale circulation anomalies and strong pressure gradients. A comparison with independent databases of European extreme events shows a statistically significant overlap, particularly for windstorms. Increasing the trajectory length enhances the detection of multi-day events, indicating that the method captures persistent atmospheric evolutions rather than isolated states. In addition to windstorms, the detected trajectories correspond to cold spells and blocking-like circulation patterns, as well as events that are not systematically documented in existing pan-European databases. These results indicate that analysing rare atmospheric trajectories provides complementary information to state-based approaches and offers a general framework for the detection of extreme atmospheric evolutions in reanalysis datasets
Overview of LifeCLEF 2025: Challenges on Species Presence Prediction and Identification, and Individual Animal Identification
No full textInternational audienceBiodiversity monitoring using AI-powered tools has become vital for tracking species distributions and assessing ecosystem health on a large scale. Automated image- and sound-based species recognition, in particular, continues to accelerate conservation efforts by enabling rapid, low-cost surveys of vulnerable populations. However, the ever-growing variety of algorithms and data sources underscores the need for standardized benchmarks to assess real-world performance. Since 2011, the LifeCLEF lab has filled this role by organizing annual evaluations that promote collaboration among AI experts, citizen science, and ecologists. In this overview, we report on the LifeCLEF 2025 edition, which featured five distinct, data-driven tasks: (i) AnimalCLEF, focusing on open-set individual animal re-identification; (ii) BirdCLEF+, about species recognition in complex acoustic soundscape recordings; (iii) FungiCLEF, addressing few-shot classification of rare fungi species; (iv) GeoLifeCLEF, combining environmental and high-resolution remote sensing with occurrence records to predict plant species presence; and (v) PlantCLEF, aiming to identify multiple co-occurring plant species in vegetation-plot imagery. This paper provides an overview of the motivation, methodology, and main outcomes of the five challenges