HAL Paris Dauphine-PSL
Not a member yet
21932 research outputs found
Sort by
La Mémorisation dans les Transformers purement attentionnels
No full textInternational audienceRecent research has explored the memoriza-tion capacity of multi-head attention, butthese findings are constrained by unrealisticlimitations on the context size. We present anovel proof for language-based Transformersthat extends the current hypothesis to anycontext size. Our approach improves uponthe state-of-the-art by achieving more effec-tive exact memorization with an attentionlayer, while also introducing the concept ofapproximate memorization of distributions.Through experimental validation, we demon-strate that our proposed bounds more accu-rately reflect the true memorization capac-ity of language models, and provide a precisecomparison with prior workDes recherches récentes ont exploré la capacité de mémorisation de l'attention à plusieurs têtes, mais ces résultats sont limités par des contraintes irréalistes sur la taille du contexte. Nous présentons une nouvelle preuve pour les Transformers qui étend l'hypothèse actuelle à n'importe quelle taille de contexte. Notre approche améliore l'état de l'art en implémentant une mémorisation exacte plus efficace avec une couche d'attention, tout en introduisant également le concept de mémorisation approximative des distributions. Par validation expérimentale, nous démontrons que les bornes que nous proposons reflètent plus fidèlement la véritable capacité de mémorisation des modèles de langage, et fournissons une comparaison précise avec les travaux antérieurs
La recherche comptable engagée : une rupture dans l’évolution de la recherche comptable ?
No full textjournée anniversaire, 30 janvier 2025 dans les locaux parisiens de NEOMA Business Schoo
Surprise sans pertinence n’est que ruine de la communication et de la créativité
No full textComment marche la créativité ? Multidimensionnelle, elle résiste à toute définition catégorique. Puissante, elle s'impose face aux turbulences du quotidien, aux aléas, aux défis contemporains. Elle engage dans la recherche de solutions, elle donne forme aux idées. Véritable force motrice, elle combine l'énergie individuelle et l'alchimie collective pour changer la donne.Ce livre clarifie le concept et communique des faits pour embrasser la richesse du sujet. Des entrepreneurs, des artistes, des formateurs, des chercheurs y témoignent d'une créativité vécue, dans des contextes variés : transformation et management, art et création, émotions et perceptions interculturelles, développement des compétences et coaching, marketing et communication, innovation et intelligence artificielle. Au fil des pages, la réflexion s'enrichit d'exemples, de 12 schémas inspirants inédits, d'outils de créativité et de carnets pratiques pour s'entrainer. Elle croise, entre autres, OpenAI et Midjourney, les Jeux Olympiques Paris 2024, les marques IKEA et Back Market, Jennifer Lopez en état de flow, les architectes Antoni Gaudi et Frank Gehry, les artistes Mark Rothko et Marina Abramovic, la Terre en forme de patate, la couleur Millennial Pink...L'ouvrage s'inscrit dans la collection "Les dessous de" qui établit des ponts entre la théorie et la pratique. Il déconstruit les idées reçues, et propose des méthodes concrètes pour manipuler les notions qui y sont exposée
SIR SQL for Logical Navigation and Calculated Attribute Free Queries to Base Tables
No full textInternational audienceSIR SQL stands for SQL with Stored and Inherited Relations (SIRs). Every SIR SQL Create Table makes definable any base attributes one could have in an SQL Create Table at present. In addition, one can define inherited attributes (IAs), definable in SQL queries or views only up to now. One may also define foreign keys (FKs) that are SQL ones or logical pointers in Codd’s original sense. IAs in SIRs with Codd’s FKs usually provide for logical navigation free (LNF) queries, i.e., without equijoins on FKs and referenced keys. The same outcome SQL queries to the same base tables without IAs, must include LN avoided.SIR SQL Create Table may in particular include IAs definable through value expressions also possible in SQL queries or views only up to now, usually referred to as calculated attributes (CAs). CAs may involve, e.g., attributes from different tables or aggregate functions, or sub-queries. CAs in SIRs provide for CAF queries, addressing any CAs in SIRs by name only. In contrast, every SQL query to base tables needing CAs has to fully define each of these. The end result is that most of SQL base table queries, requiring LN or CAs schemes at present, become LNF or CAF queries in SIR SQL. The latter are usually substantially less procedural, i.e., by dozens of characters. They become also quasi-natural, i.e., with Select clause only naming the selected attributes, From clause naming a single base table and Where clause, if any, with short Boolean formulae over usual constraints on some attribute values, at worst. SIR SQL should accordingly significantly boost SQL clients’ productivity. Especially, since most clients are data analysts or application developers, not SQL geeks. While the problematic of LNF and CAF queries is four decades old, our solution is the first practical one, to our best knowledge.Below, we illustrate the problem of LN and of CAs in queries to SQL base tables using Codd’s original Supplier-Part DB. We then present SIR SQL. We show in depth how SIR SQL LNF and CAF queries to base tables become possible. We show in particular that Create Table statements defining an SQL DB at present, usually define also a SIR SQL DB, providing for LNF queries to base tables as free bonus. We discuss the front-end for SIR SQL that should require, for any popular SQL DBS, a few month implementation efforts only, validated by proof-of-concept prototype for SQLite3. We accordingly postulate to upgrade every popular SQL DBS to SIR SQL. 7+ million SQL clients worldwide, of the dominant DB language, providing for 31B+ US$ market size of SQL apps, will benefit from
LP-4P : Link Prediction over Annotated Knowledge Graphs with Four Patterns
No full textInternational audienceInitially designed for Knowledge Graphs (KG) containing plain (s, p, o) triples, research on link prediction has evolved to address more complex structures such as hypergraphs, hyper-relational graphs, and bi-level graphs. These advancements push link prediction to handle annotation patterns where a triple is annotated with another (t, p, t), or where a triple is annotated with qualifiers (t, p, o). However, each approach focuses on a single annotation pattern, and the case where a triple is annotated with a subject (s, p, t) has never been explored. In this paper, we propose LP4P, the first link prediction model capable of predicting entities in both plain triples and three annotation patterns. LP4P captures the information expressed in annotations through a four-pattern attention layer, and its loss function further leverages the ontological information of KGs. Moreover, we built WD4P, an RDF KG derived from baseline datasets, which includes the four patterns. Extensive empirical evaluations demonstrate that LP4P outperforms relevant state-of-the-art models on knowledge graphs with the four patterns, and achieves comparable performance on standard benchmarks
Exact output tracking for the one-dimensional heat equation and applications to the interpolation problem in Gevrey classes of order 2
No full textThis paper provides a complete characterization of the Dirichlet boundary outputs that can be exactly tracked in the one-dimensional heat equation with Neumann boundary control. The problem consists in describing the set of boundary traces generated by square-integrable controls over a finite or infinite time horizon. We show that these outputs form a precise functional space related to Gevrey regularity of order 2. In the infinite-time case, the trackable outputs are precisely those functions whose successive derivatives satisfy a weighted summability condition, which corresponds to specific Gevrey classes. For finite-time horizons, an additional compatibility condition involving the reachable space of the system provides a full characterization. The analysis relies on Fourier-Laplace transform, properties of Hardy spaces, the flatness method, and a new Planchereltype theorem for Hilbert spaces of Gevrey functions. Beyond control theory, our results yield an optimal solution to the classical interpolation problem in Gevrey-2 classes, which improves results of Mitjagin on the optimal loss factor. The techniques developed here also extend to variants of the heat system with different boundary conditions or observation points
Dépasser la guerre contre les microbes
No full textLa semaine mondiale de sensibilisation à l’antibiorésistance se termine. Sommes-nous sur la voie de résoudre ce problème qui risque, d’ici quelques dizaines d’années, de rendre mortelles des infections banales ? Il ne suffit pas de dénoncer la mauvaise utilisation individuelle par les patients, médecins et vétérinaires. Il faudrait plutôt interroger la dépendance aux antibiotiques de nos systèmes sanitaires et agro-alimentaires. Et même considérer autrement notre collaboration avec le vivant
La résilience du productivisme à l’heure de l’antibiodépendance
No full textInternational audienceLes efforts pour limiter les antibiotiques en élevage se multiplient, mais le modèle industriel persiste. La question de fond est rarement abordée : comment repenser un système agroalimentaire structuré pour la productivité plutôt que pour la durabilité
Goodness-of-fit testing for the stationary density of a size-structured PDE
No full textWe consider two division models for structured cell populations, where cells can grow, age and divide. These models have been introduced in the literature under the denomination of 'mitosis' and 'adder' models. In the recent years, there has been an increasing interest in biology to understand whether the cells divide equally or not, as this can be related to important mechanisms in cellular aging or recovery. We are therefore interested in testing the null hypothesis where the division of a mother cell results into two daughters of equal size, against the alternative hypothesis where the division is asymmetric and ruled by a kernel that is absolutely continuous with respect to the Lebesgue measure. The sample consists of i.i.d. observations of cell sizes and ages drawn from the population, and the division is not directly observed. The hypotheses of the test are reformulated as hypotheses on the stationary size and age distributions of the models, which we assume are also the distributions of the observations. We propose a goodness-of-fit test that we study numerically on simulated data before applying it on real data
Bridging the Theoretical Gap in Randomized Smoothing
No full textInternational audienceRandomized smoothing has become a leading approach for certifying adversarial robustness in machine learning models. However, a persistent gap remains between theoretical certified robustness and empirical robustness accuracy. This paper introduces a new framework that bridges this gap by leveraging Lipschitz continuity for certification and proposing a novel, less conservative method for computing confidence intervals in randomized smoothing. Our approach tightens the bounds of certified robustness, offering a more accurate reflection of model robustness in practice. Through rigorous experimentation we show that our method improves the robust accuracy, compressing the gap between empirical findings and previous theoretical results. We argue that investigating local Lipschitz constants and designing adhoc confidence intervals can further enhance the performance of randomized smoothing. These results pave the way for a deeper understanding of the relationship between Lipschitz continuity and certified robustness