Portail HAL Lumière Lyon 2
Not a member yet
111406 research outputs found
Sort by
Impact of physiological and biomechanical parameters on lung deformation and the accuracy of lung tumor motion estimation
No full textInternational audiencePatient-specific biomechanical models of the respiratory system can enhancethe prediction of lung tumor positions and deformations for radiation ther-apy. To achieve this, we have developed a patient-specific biomechanicalmodel of the entire respiratory system. However, the accuracy of the simula-tion is highly influenced by mechanical behavior as well as biomechanical andphysiological properties. In this study, we have investigated the impact ofsimplification and variability in mechanical and physiological property uncer-tainties on lung tumor motion prediction. Specifically, we have evaluated andcompared the most commonly used values of the lung tissue Young’s modulusand Poisson’s ratio found in the literature. Furthermore, we have examinedthe effect of a simple and fast linear compliance model versus a nonlinear,personalized physiological lung compliance model in computing lung and di-aphragm strain. We have also explored the impact of different nonlinearbehavior models to identify the most suitable mechanical model for respi-ratory simulation. To this end, we have conducted a study on four widelyreferenced hyperelastic models. Numerical simulations were performed onpublic datasets using the Neo-Hooke, Yeoh, Mooney-Rivlin, and St. Venant-Kirchhoff hyperelastic models. We have observed that nonlinear personalizedcompliance enhances accuracy and yields better results compared to linearcompliance. The simulations in this study showed minimal and negligiblevariations with different values of Young’s modulus. In contrast, variationsin Poisson’s ratio significantly impacted the simulation results. In our simula-tions, the Saint-Venant–Kirchhoff and Mooney–Rivlin models demonstratedthe highest accuracy for simulating lung tissue across all phases of respira-tion, with an average landmark error of 2.1 ± 1.3mm. This model has the potential to provide precise tumor motion predictions, helping physicians re-duce safety margins and minimize damage to healthy tissues during radiationtherapy
Idiom learning: conceptual metaphors as a metacognitive tool in the EFL classroom
No full textThis master’s thesis aims at applying the theory of conceptual metaphors developed by linguists George Lakoff and Mark Johnson in Metaphors We Live By (1980), with high school students in the EFL classroom, in order to boost their learning and retention of metaphorical idioms. The analysis of conceptual metaphors conducted in a first part allows me to devise a metaphor-based teaching method. This method is then implemented in a two-hour workshop dispensed to Terminale students with Contemporary World English (CWE) specialisation (Anglais Monde Contemporain). Finally, in lights of the execution of the workshop as well as the students’ feedback, I analyse the method.Ce mémoire vise à appliquer la théorie des métaphores conceptuelles développée par les linguistes George Lakoff et Mark Johnson dans leur ouvrage Metaphors We Live By (1980), avec des élèves de lycée en classe d’anglais langue étrangère, afin d’optimiser leur apprentissage et rétention des expressions idiomatiques métaphoriques. L’analyse de cette théorie menée dans un premier temps permet d’élaborer une méthodologie centrée sur les métaphores conceptuelles. Dans un deuxième temps, cette méthodologie est mise en œuvre à travers une séance de deux heures dispensées à des élèves de Terminale en enseignement de spécialité Anglais Monde Contemporain (AMC). Enfin, une évaluation du déroulé de la séance et du retour d’expérience des élèves permet d’analyser la méthodologie
Modeling Energy Consumption in Deep Learning Architectures Using Power Laws
No full textInternational audienceModern Deep Learning architectures such as LSTM, GRU, and Transformers achieve remarkable performance in various sequence processing tasks. Yet, their high computational cost and energy consumption have raised concerns about their environmental impact and the sustainability of Deep Learning. In this paper, we present an empirical study assessing the efficiency of training LSTM, GRU, and Transformer models on a GPU. By evaluating these models under various configurations, we characterize the relationship between energy consumption and pre-defined quantities such as hardware efficiency and the number of floating point operations (FLOPs) required for inference. We show that it is possible to derive scaling laws that make energy consumption predictable, given an architecture and a GPU model.</div
Urban neo-feudalism as the hidden face of extended urbanization? On the « ruralisation » of urban regulation in a postcolonial capital
No full textInternational audienc
La rébellion stérile des géniteurs féconds face à l’expertise génétique
No full textInternational audienc
Accepting to incarcerate in immediate trial. How judges share otherizing considerations of defendants in backstage
No full textInternational audienceIn this communication I examine the subjective relationship that judges have with the possibility of inflicting suffering by incarcerating defendants in the french procedure of immediate trials. I first demonstrate the need for judges to distance themselves from an understanding and empathetic stance towards defendants. Secondly, I inform how the professional environment encourages mechanisms for neutralizing compassion
Apprentissage incrémental de l'opérateur de Koopman pour systèmes non-autonomes via prédiction conforme
No full textInternational audienceThe Koopman operator enables the linearization of certain nonlinear dynamical systems by acting on observables, but its extension to the non-autonomous setting remains challenging. We propose an incremental method to estimate the Koopman operator in a non-autonomous context, leveraging conformal prediction to automatically detect regime changes. Model parameters are updated only when local performance deteriorates, ensuring sparse updates while maintaining accuracy comparable to off-line training. Preliminary experiments on the Duffing oscillator illustrate the effectiveness of our approach.L'opérateur de Koopman permet de linéariser certaines dynamiques non linéaires via une action sur les observables, mais son extension au cas non-autonome reste un défi. Nous proposons une méthode incrémentale pour estimer cet opérateur dans un cadre non-autonome, en s'appuyant sur la prédiction conforme pour détecter automatiquement les changements de régime. Les paramètres du modèle sont ajustés uniquement lorsque la performance locale se dégrade, assurant des mises à jour parcimonieuses tout en conservant une précision comparable à l'apprentissage hors-ligne. Des expériences préliminaires sur le système de Duffing illustrent l'efficacité de notre approche
Lithic industries, modern tools, and language: An evolutionary perspective through fMRI
No full textInternational audienceTool use is a defining cognitive ability of the human species, relying on technical reasoning-a causal and analogical understanding of physical principles. Neuroarchaeological studies suggest that lithic tool use engaged a specialized frontoparietal network, which evolved into the left-lateralized network observed for modern tools. This network includes the left inferior parietal lobule, particularly the area PF, linked to technical reasoning, and the left inferior frontal gyrus, which is both involved in tool use and language. Since these latter two domains are based on hierarchical structuring, organizing embedded constraints, the common involvement of the left inferior frontal gyrus has been proposed as evidence of a co-evolutionary trajectory. Using fMRI, we investigated whether increasing mechanical complexity modulates frontoparietal activity and whether lithic and modern tools engage a common neural network. Participants performed a tool evaluation task, in which they assessed the functionality of both tool types across three levels of complexity designed to reflect embedded constraints. Our results revealed stronger functional connectivity between the left area PF and the left pars opercularis of the inferior frontal gyrus as mechanical complexity increased. The results also confirmed common activation for both tool types. By demonstrating that frontoparietal connectivity scales with complexity in tool evaluation, our study provides new insights into the neurocognitive foundations of tool use. These findings contribute to the broader discussion of a co-evolutionary relationship involving technical reasoning, tool making, and language, highlighting the role of hierarchical processing as a potentially shared computational principle
Robotic Manipulation via Imitation Learning: Taxonomy, Evolution, Benchmark, and Challenges
No full textRobotic manipulation (RM) is central to enabling autonomous robots to interact with and alter their environments in real-world scenarios. Among the learning paradigms, imitation learning has emerged as a powerful approach, allowing robots to rapidly acquire complex manipulation skills from human demonstrations. This survey provides the first systematic review dedicated to imitation learning for robotic manipulation. We identify and analyze a large set of representative studies selected for their scientific quality and community impact. For each, we provide a structured summary covering purpose, technical implementation, taxonomy, input formats, priors, strengths, limitations, and citation metrics. Beyond cataloging, we trace the chronological evolution of imitation learning techniques within robotic manipulation policies (RMPs), highlighting key methodological shifts-from diffusion and flow matching to autoregressive and affordance-driven strategies. Where available, we compile benchmark results and conduct quantitative comparisons, enabling an integrated view of performance across tasks and environments. Finally, we outline open challenges such as generalization, embodiment diversity, data efficiency, and benchmark standardization, and we discuss promising directions toward scalable and general-purpose robotic manipulation. By synthesizing methods, benchmarks, and challenges, this survey aims to serve both as an entry point for newcomers and a reference for active researchers seeking to advance imitation learning in robotic manipulation
The Wheat and Rice Genomics Scientific Literature Knowledge Graphs
No full textInternational audienceThis paper presents a generic semantic model to describe, structure, and integrate the named entities automatically extracted from scientific texts, represented as annotations. This model has been used to construct knowledge graphs from two distinct agricultural corpora consisting of PubMed scientific publications on wheat and rice genetics. The named entities to be recognized are genes, phenotypes, traits, genetic markers, and taxa. For both corpora, named entities were automatically extracted using natural language processing tools. The RDF model was populated using a mapping-based transformation pipeline implemented with the Morph-xR2RML tool which takes CSV files as input. The resulting RDF knowledge graphs are deployed and query-able through dedicated web applications