HAL Paris Dauphine-PSL
Not a member yet
21932 research outputs found
Sort by
Memorization vs. Generalization in diffusion models with the U-Net architecture: A Reproduction with Perceptual Metrics and Attribute-Controlled Splits
No full textInternational audienceThis report reproduces and extends the experiments of Kadkhodaie, Guth, Simoncelli and Mallat (2024) on generalization in diffusion models. The original study reported that two U-Net denoisers trained independently on non-overlapping subsets of CelebA produce nearly identical samples when seeded with the same noise, suggesting a form of deterministic convergence. We confirm this qualitative behavior under constrained settings (40×40 resolution) and develop a fully reproducible, memory-efficient preprocessing pipeline. Beyond replication, we explore its stability with respect to dataset correlation by introducing attribute-controlled (eyeglasses/no-eyeglasses, male/female) and identity-disjoint splits, and by assessing sample originality with a perceptual metric (LPIPS) instead of pixel correlation. Our findings indicate that convergence between independently trained denoisers remains strong when training subsets are correlated, but weakens as semantic heterogeneity increases. These observations suggest that both architectural inductive biases and dataset redundancy contribute to the stability of the phenomenon originally reported by Mallat and collaborators
Does rent control increase rental returns ? The case of the metropolitan housing market of Lille, France
No full textUsing machine-learning (ML) methods to appraise rental returns with a rent and a transaction database, we estimate a difference-in-differences (DiD) model to test the effect of rent control (RC) on rents, capital returns and rental returns in Lille, France. Bringing a financial perspective to public policy analysis, we show that the introduction of RC, a tool that is the subject of much debate in economic literature, has negative consequences on the housing market through the signal it sends out to investors rather than through its direct effects. It has led to i) an increase in rents for the largest properties, ii) a decrease in capital returns for all properties and iii) an increase in rental returns for the largest ones. These results are consistent with the predictions of the stock-flow model considering the effects of investor expectations. RC has led to a decrease in housing prices rather than a general decline in rents or rental returns, especially for the largest properties
A critical genealogy of time control devices in France
No full textThis chapter analyses the non-linear evolution of working time control systems, focusing on their role in the management and control of working time in France. The decision to focus on France is motivated by the extreme uniqueness of certain employment contracts, which makes it difficult and irrelevant to compare them with other countries. Despite this bias, it will be necessary to look at the development of techniques developed in the United States in particular since the use of management tools is both highly specific (motivated by the adoption of a law, for example) and universal (spread of a managerial ideology, international character of companies such as IBM or technologies). The assumption of a linear progression of these tools is problematic as it underestimates the influence of social, economic, and political factors in their adoption and use. This study is essential to better understand how time control devices (TCDs) shape the structure of organisations, particularly for managers, and how they influence power dynamics within companies. Critical analysis of TCDs can offer ways of re-evaluating assumptions often taken for granted in studies of management and organisational history. Moving away from simplistic narratives of linear technological progress, we explore the complex interplay between time control devices and broader societal factors. Our analysis highlights the nuanced ways in which these devices shaped and were shaped by French economic, social, and cultural conditions. While not exhaustive, our focus on punch cards and timesheets provides a rich case study of how time control devices were used to regulate labour, shape power dynamics, and influence organisational structures. It also offers an alternative to the view that technological innovation is the driver of change
« Les enseignants-chercheurs français en gestion face aux nouvelles injonctions institutionnelles : Consentement, résistance ou servitude volontaire ? »
No full textLe mot « consentement » est entré aujourd’hui dans la conversation courante, mais essentiellement sous l’angle de rapports sexuels non consentis. Céder n’est pas consentir. De nombreux livres témoignages ont été publiés. Ce livre déborde la sphère de l’intime et aborde la sphère publique pour poser aujourd’hui une question clé : consentir à quoi ? À cette question, cet ouvrage collectif tiré du colloque de l’École supérieure de commerce de Paris (ESCP Business School) « Consentir ? Pourquoi, comment et à quoi ? » y répond en utilisant plusieurs plans d’analyse : économique, juridique, psychologique, sociologique, politique, anthropologique. Faire appel aux ressources apportées par les sciences de l’homme et de la société permet de déployer l’ambivalence de la notion de consentement. Consentir suppose une relation contractuelle conclue entre égaux. Consentir à payer des impôts relève-t-il d’une adhésion volontaire ? Cocher la case « consentir » pour avoir accès à une base de données relève-t-il d’une simple procédure ? Comment réagissons-nous aujourd’hui face au célèbre problème de la servitude volontaire mis en lumière en son temps par La Boétie ? Pourquoi sommes-nous disposés à collaborer à notre propre exploitation lorsque nous travaillons ? Sommes-nous à ce point aveugles pour ne pas voir que le consentement se fabrique sous nos yeux, à l’insu de notre plein gré
Action collective créatrice, management et interculturalité : une perspective anthropologique
No full textCet ouvrage, issu d'un colloque de Cerisy, est le premier à explorer l'ensemble des travaux d'Armand Hatchuel pour analyser son itinéraire théorique et les multiples implications de ses propositions. Des chercheurs venant de plusieurs disciplines (sciences de gestion, droit, philosophie, économie, histoire…), ainsi que plusieurs dirigeants d'entreprise, situent la place que ces travaux ont eu dans leurs disciplines ou pour leurs propres parcours. Certes, les travaux d'Armand Hatchuel soulignent les illusions d'une « économie politique », d'une « science sociale » ou de « sciences de gestion » restées prisonnières des dogmes de l'agir moderne. Mais surtout, ils jettent les bases rigoureuses d'une « science de l'action collective créatrice » comme socle indispensable pour « bien gérer » dans l'anthropocène. Un projet désormais vital
NONLINEAR FLUCTUATIONS FOR A CHAIN OF WEAKLY ANHARMONIC OSCILLATORS WITH STOCHASTIC PERTURBATION
No full textWe study the fluctuations of the phonon modes in a one-dimensional chain of anharmonic oscillators where the deterministic Hamiltonian dynamics is perturbed by random exchanges of momentum between nearest neighbor particles. There are three locally conserved quantities: volume, momentum and energy. We study the evolution in equilibrium of the fluctuation fields of the two phonon modes (linear combination of the volume stretch and momentum), on a diffusive space-time scale after recentering on their sound velocities. We show that, weakening the anharmonicity with the scale parameter, the recentered phonon fluctuations fields converge to the stationary solutions of two uncoupled stochastic Burgers equations. The nonlinearity in the Burgers equation depends on the presence of a cubic term in the anharmonic potential (corresponding to the α-FPUT dynamics). Main ingredients of the proof, based on a compactness argument for the Dynkin's martingale decomposition, are the second-order Boltzmann-Gibbs principle, as well as equipartition of energy, to characterize the nonlinear term and a Riemann-Lebesgue estimates showing that fields with diverging velocity to different directions have no interaction in the limit.</div
Physician knowledge, attitudes, and perceptions of respiratory syncytial virus in older adults: A cross-sectional survey in Germany and Italy
No full textInternational audienceRespiratory syncytial virus (RSV) causes substantial morbidity and mortality among older adults and individuals with chronic diseases, putting pressure on healthcare systems. RSV vaccines became available in 2023, however, physicians’ knowledge and attitudes regarding RSV and RSV vaccines may hinder uptake. We aimed to understand physicians’ knowledge, attitudes, and perceptions of respiratory infections, including RSV, and vaccination. A pre-tested cross-sectional survey was fielded with physicians who commonly administer or recommend respiratory vaccines in Germany and Italy. Regression models assessed physicians’ characteristics associated with knowledge and perceptions of RSV. Overall, 307 physicians (Germany: n = 152; Italy: n = 155) completed the survey. Physicians had good mean RSV disease knowledge scores (Germany: 3.8/5; Italy: 3.3/5). Most (Germany: 68%; Italy: 72%) indicated wanting further information on RSV, versus <35% for other respiratory diseases. Most considered RSV an important burden, especially for adults aged ≥50 years with immunocompromising conditions (Germany: >96%; Italy: >92%) and ≥60 years with chronic conditions (Germany: >95%; Italy: >87%). Physicians seeing >100 versus ≤100 patients weekly, and general practitioners versus other specialists, perceived RSV burden as more important. Perceived barriers to RSV vaccination included lack of national recommendation and reimbursement (Germany: 92%; Italy: 89%), and patients not being informed on getting vaccinated (Germany: 71%; Italy: 59%). In conclusion, physicians generally considered RSV an important pathogen for older adults and those with chronic conditions at risk of severe infection. Most wanted more information on RSV, thus medical education is important to address knowledge gaps and enable physicians to guide patients in making informed vaccination decisions
Apprentissage profond basé sur les modèles pour la gestion du risque cyber
No full textIn this thesis, we study the potential of hybrid learning methods for predicting intrusion "kill chains" within an enterprise grade network. For a computer network represented by a directed graph, we seek an algorithm capable of predicting the path taken by the attacker in three configurations: (a) for a given entry point (known as the "initial access") and exit point (known as the "impact point"); (b, c) when only one of these points is known. Beyond graph traversal, and in order to differentiate ourselves from existing methods, we seek to determine whether it is possible to take into account - via a learning algorithm - a dataset of cyber threat intelligence, integrated in the form of a knowledge graph. Due to the scarcity of datasets describing both the compromised environment and the path taken by the attacker, we propose to experiment with the use of hybrid methods from the signal processing and applied physics communities, known as "Model-Based Deep Learning" or "Physics-Informed Neural Networks". These architectures aim to guide the model - through the loss function- according to a set of structural properties known about the process under study. Our results show that this approach improves generalisation compared to traditional learning. Finally, we propose a complete software implementation of the models developed, as part of a risk management process aligned with the ISO/IEC-27005 standard.Dans cette thèse, nous étudions le potentiel des méthodes d'apprentissage hybrides pour la prédiction de chaînes d'intrusion informatique au sein d'un réseau d'entreprise. Pour un réseau informatique représenté par un graphe dirigé, nous cherchons un algorithme capable de prédire le chemin parcouru par l'attaquant dans trois configurations : (a) pour un point d'entrée (dit « accès initial ») et un point de sortie (dit « point d'impact ») donnés; (b, c) lorsque seul l’un de ces points est connu. Au-delà du parcours de graphe, et afin de nous différencier des méthodes existantes, nous cherchons à déterminer s’il est possible de prendre en compte - via un mécanisme d'apprentissage - un ensemble de renseignements sur les cybermenaces, intégré sous la forme d'un graphe de connaissances. En raison de la rareté des jeux de données décrivant à la fois l'environnement compromis et le chemin emprunté par l'attaquant, nous proposons d’expérimenter l’utilisation de méthodes hybrides issues des communautés du traitement du signal et de la physique appliquée, appelées « Apprentissage Profond Basé sur les Modèles » ou « Apprentissage Informé par la Physique ». Plus frugales, ces architectures visent à guider le modèle - à travers la fonction de perte - selon un ensemble de propriétés structurelles connues du processus étudié. Nos résultats permettent de conclure à une amélioration des performances grâce à cette méthode, par rapport à un apprentissage classique. Enfin, nous proposons une implémentation logicielle complète des modèles développés, s’inscrivant dans un processus de gestion des risques aligné sur le standard ISO/IEC-27005
Analyser les expériences d’injustice et de discriminations à l’école : contextualisation, ressources et ambivalences: Apports de méthodes mixtes
No full textInternational audienceBased on a mixed-method study – a questionnaire survey and a qualitative post-survey (TeO1-Ined-Insee) – this article examines the processes and reasons that lead actors to perceive a situation as discriminatory in the school environment, even though the principle of academic meritocracy is a shared value in the French context. Three elements form the basis for these claims of discrimination: identity dissonance, family context and educational segregation, which nevertheless has ambivalent effects on the analysis that actors make of their experiences of discrimination. This article demonstrates the value of mixed methods for measuring, understanding and shedding light on a sensitive and complex subject: discrimination experienced at school.À partir d’une étude à méthodologie mixte – une enquête par questionnaire et une post-enquête qualitative (TeO1-Ined-Insee), cet article interroge les processus et les raisons qui amènent les acteurs à percevoir une situation comme discriminante dans l’univers scolaire, alors que le principe de méritocratie scolaire constitue une valeur partagée dans le contexte français. Trois éléments constituent des points d’appui dans ces déclarations de discrimination : les dissonances identitaires, le contexte familial et la ségrégation scolaire qui, pour ce dernier, a néanmoins des effets ambivalents sur l’analyse que les acteurs font de leurs expériences de discrimination. Cet article montre l’intérêt des méthodes mixtes pour mesurer, appréhender et éclairer un objet sensible et complexe : celui des discriminations vécues à l’école