Scientific Publications of the University of Toulouse II Le Mirail
Not a member yet
92205 research outputs found
Sort by
Busemann Functions in the Wasserstein Space: Existence, Closed-Forms, and Applications to Slicing
No full textThe Busemann function has recently found much interest in a variety of geometric machine learning problems, as it naturally defines projections onto geodesic rays of Riemannian manifolds and generalizes the notion of hyperplanes. As several sources of data can be conveniently modeled as probability distributions, it is natural to study this function in the Wasserstein space, which carries a rich formal Riemannian structure induced by Optimal Transport metrics. In this work, we investigate the existence and computation of Busemann functions in Wasserstein space, which admits geodesic rays. We establish closed-form expressions in two important cases: one-dimensional distributions and Gaussian measures. These results enable explicit projection schemes for probability distributions on , which in turn allow us to define novel Sliced-Wasserstein distances over Gaussian mixtures and labeled datasets. We demonstrate the efficiency of those original schemes on synthetic datasets as well as transfer learning problems
Combining word and phoneme speech recognition for fluency assessment of young children's oral reading
No full textInternational audienceFluency is a key component of learning how to read, as it unlocks comprehension skills. However, it is not easy for teachers to monitor their students' fluency practice in class. With the ASR fluency assessment system presented in this work, we aim at enabling students to improve their reading skills through practice and helping teachers assess their progress. Since fluency assessment is still rather an uncommon task, we start by defining the feedback expected by teachers and establish a set of metrics to evaluate our automatic system's performance. We compare individual phoneme-level and wordlevel systems and design an innovative system that combines them. This combined system obtains the best WCPM MAE, accuracy MAE and F1-score, showing that it leverages multilevel knowledge for a more accurate detection of reading mistakes in students' oral reading. We also discuss the behavior of our system depending on the student's age and demographic information to detect biases to reduce.
Hybrid extended cavity laser made of silicon nitride Bragg gratings and GaAs optical amplifiers for frequency comb generation around 965 nm
No full textInternational audienceWe present the characterisation of a silicon nitride chip that includes a Bragg reflector, which, butt-coupled to a Reflective Semiconductors Optical Amplifier, will form an extended cavity laser emitting around 965nm for generating a frequency comb
Analysis of quasi-periodic waves of cubic nonlinear Schrödinger equations
No full textWe study the quasi-periodic standing wave solutions of the focusing and defocusing cubic nonlinear Schrödinger equations in dimension one. In the defocusing case, we establish a diffeomorphic correspondence between the invariants of the ordinary differential equation of the wave profiles and the conserved quantities of the evolution equation. We introduce a numerical scheme to compute the minimizers of the energy at fixed mass and momentum for both focusing and defocusing cases. The scheme is based on a gradient flow approach with discrete renormalization at each time step. The novelty of our scheme is that the renormalization step deals at the same time with the mass and the momentum constraints. In numerical experiments, we observe that a given solution of the profile ordinary differential equation is also a minimizer of the energy at corresponding mass and momentum
Study of temperature-dependent breakdown in AlGaN/GaN normally-off HEMT under drain step-stress
No full text#45 Final article depositen collaboration avec le LAAS-CNRS, l'université de Toulouse, le CNES Toulouse et l'ADERA.International audienceThis paper investigates the temperature-dependent breakdown of commercially available GaN-on-Si power transistorsduring off-state drain step-stress. This setup allows us to obtain novel results on commercially available 650V p-GaN HEMTsubjected to step-stress under various conditions of temperature and gate voltage. Previously observed breakdown mechanismsare compared with the one observed on the current structure and a physical interpretation is proposed to explain thephenomenon. Activation energy, Weibull parameters and failure analysis are explored to provide understanding on the failuremechanisms
Democracy in organisations as a prerequisite for the vitality of democracy in society. The special role of digital information systems.
No full textInternational audienceIn our so-called liberal societies, citizens are only part-time citizens: democracy ends at the doors of the organisations in which they work. Our fellow citizens therefore lead double lives: they are citizens who are called upon to express their opinions in public life, but they are also employees who must submit to the decision-making power of top management (or even shareholders alone).We argue that this dichotomy is detrimental to the health of our democracies.Furthermore, it jeopardises the chances of achieving a genuine ecological transition within organisations, which requires genuine eco-determination.There are many obstacles, but also some levers, to the advent of democracy within organisations. This presentation aims to analyse the role of digital information systems (IS) in organisational democracy, or indeed in its very possibility.IS within organisations are often presented as neutral technical tools. We will demonstrate that IS are political instruments in that they do not simply reflect reality, but actively contribute to its construction through the representations that the organisation embodies. Consequently, decisions are often made within the limited and constrained framework of the digitised reality of IS.Furthermore, IS impact employees' autonomy, as well as on their working rhythm and conditions, with little opportunity for them to influence the IS design. The IS are political instruments that are not democratically built.In conclusion, we emphasise the importance of co-designing these systems through a sustained democratic dialogue between technical experts, managers, and future users.Dans nos sociétés dites libérales, les citoyens ne le sont qu'à temps partiel : la démocratie s'arrête en effet aux portes des organisations dans les lesquelles ils travaillent. Nos concitoyens mènent donc une double vie : citoyens appelés à exprimer leurs opinions dans la vie de la cité, et salariés qui doivent se soumettre au pouvoir décisionnaire du haut management (voire des seuls actionnaires).Notre position est que cette dichotomie est préjudiciable à la bonne santé de nos démocraties. Par ailleurs, elle compromet les chances que soit menée une véritable transition écologique au sein des organisations, laquelle nécessite une véritable écodétermination.De nombreux obstacles, mais aussi quelques leviers, existent à l'avènement de la démocratie au sein des organisations. L'objet de cette présentation est d'analyser plus particulièrement le rôle des systèmes d'information numériques (SI) dans la démocratie organisationnelle (ou dans sa possibilité même).Les SI des organisations sont souvent présentés comme de simples outils techniques, neutres. Nous montrerons que les SI des organisations sont des dispositifs politiques, au sens où ils ne codent pas de façon neutre une réalité immanente, mais au contraire participent à construire cette réalité au travers des représentations dont l'organisation est porteuse. Les décisions sont ensuite, le plus souvent, prises dans le cadre limité et contraignant de la réalité numérisée dans les SI. Nous conclurons en insistant sur la nécessité que ces systèmes soient co-conçus, en menant un dialogue démocratique soutenu, réunissant experts technique et futurs utilisateurs
Vers une approche basée sur les graphes de connaissances pour l'évaluation de la qualité des données dans l'IoT
No full textInternational audienceL'évaluation de la qualité des données (QD) dans les systèmes IoT constitue un défi en raison de l'hétérogénéité des données, de leur production en flux continus et des environnements dynamiques dans lesquels elles sont produites. Les approches existantes se limitent souvent à un nombre restreint d'indicateurs de qualité. Cet article propose une approche granulaire et traçable reposant sur des graphes de connaissances (GC) pour enrichir et annoter les données avec des informations détaillées sur leur qualité. Celle-ci prend en compte divers indicateurs en fonction du contexte de production des données et de la tâche cible. Trois niveaux de granularité sont définis afin de structurer l'évaluation des indicateurs de QD, et des ontologies dédiées à cette évaluation sont présentées dans le cadre de ce travail
Cascade Computation of Second Order Information for Efficient Optimization of Multi-Argument Objective Functionals
No full textInternational audienceWe present a general technique for efficient computation of second order information for optimization functionals involving many variables of different nature, frequently appearing in modern applications. This technique relies on the bilinear Hessian, and while the main tools for this task have been published elsewhere, we here focus on the problem of how to organize the computations in the multi-argument situation, relying on so called cascade optimization and the chain rule for the bilinear Hessian. For concreteness, we demonstrate our approach by focusing on three concrete problems; phase retrieval, multidistance nano-holotomography and training of neural networks. In the experimental section we present results where the proposed technique significantly speeds up first order methods, and in particular we obtain faster training results than the Adamalgorithm when training the so called SIREN neural network
auto-xFS: An explanation-based feature selection tool for more meaningful and trustworthy machine learning models
No full textInternational audienceauto-xFS is a novel tool for feature selection (FS) based on a three-dimensional perspective encompassing feature retention rate, machine learning (ML) model performance, and explainability (XML). The application is designed to streamline the user’s workflow by autonomously hyperparameterizing FS techniques, ML models, and XML methods using a meta-learning approach. Our findings demonstrate that prioritizing FS that yields highly precise prediction explanations even at the expense of a slight reduction in model accuracy, can ensure more meaningful information for the user. auto-xFS is totally suited to FS in critical areas such as biomedicine where user confidence is crucial
Development and Feasibility of a Mixed Virtual and in-Presence Therapeutic Education Program for Early Multiple System Atrophy Patients
No full textInternational audienceBACKGROUND: Multiple system atrophy (MSA) is a relentlessly progressive neurodegenerative disorder with no cure. While therapeutic education programs (TEPs) benefit many chronic conditions, data in MSA are limited. OBJECTIVES: To assess the feasibility and satisfaction of a mixed in-person/virtual TEP for early-stage MSA patients and caregivers. METHODS: A six-module, self-paced TEP was co-developed over 2 years by MSA healthcare professionals, patients, and education experts. A prospective pilot study enrolled recently diagnosed MSA patients (≤12 months) and caregivers. Primary outcomes were recruitment and retention rates. Secondary outcomes included disease progression (UMSARS I-II), health-related quality of life (Hr-QoL), compared to matched controls not enrolled in the TEP and patients' satisfaction. RESULTS: Fifteen patient-caregiver pairs were enrolled. Recruitment and retention rates were 44% and 86%, respectively; 78% of modules were completed (least consulted was disease progression). Disease severity and Hr-QoL worsened over 12 months, with no significant differences versus controls. Overall, 90% of participants were satisfied or very satisfied. CONCLUSIONS: Our pilot study shows that a mixed TEP for MSA patients and familial caregivers is feasible and able to engage them over time