Alessandro Leite

Un utilisateur centré et multi-cloud architecture pour le calcul des applications de haute performance: Le cloud computing a été considéré comme une option pour exécuter des applications de calcul haute performance. Bien que les plateformes traditionnelles de calcul haute performance telles que les grilles et les supercalculateurs offrent un environnement stable du point de vue des défaillances, des performances, et de la taille des ressources, le cloud computing offre des ressources à la demande, généralement avec des performances imprévisibles mais à des coûts financiers abordables. Pour surmonter les limites d’un cloud individuel, plusieurs clouds peuvent être combinés pour former une fédération de clouds, souvent avec des coûts supplémentaires légers pour les utilisateurs. Une fédération de clouds peut aider autant les fournisseurs que les utilisateurs à atteindre leurs objectifs tels la réduction du temps d’exécution, la minimisation des coûts, l’augmentation de la disponibilité, la réduction de la consommation d’énergie, pour ne citer que ceux-Là. Ainsi, la fédération de clouds peut être une solution élégante pour éviter le sur-Approvisionnement, réduisant ainsi les coûts d’exploitation en situation de charge moyenne, et en supprimant des ressources qui, autrement, resteraient inutilisées et gaspilleraient ainsi de énergie. Cependant, la fédération de clouds élargit la gamme des ressources disponibles. En conséquence, pour les utilisateurs, des compétences en cloud computing ou en administration système sont nécessaires, ainsi qu’un temps d’apprentissage considérable pour maîtrises les options disponibles. Dans ce contexte, certaines questions se posent: (a) Quelle ressource du cloud est appropriée pour une application donnée? (b) Comment les utilisateurs peuvent-Ils exécuter leurs applications HPC avec un rendement acceptable et des coûts financiers abordables, sans avoir à reconfigurer les applications pour répondre aux normes et contraintes du cloud ? (c) Comment les non-Spécialistes du cloud peuvent-Ils maximiser l’usage des caractéristiques du cloud, sans être liés au fournisseur du cloud ? et (d) Comment les fournisseurs de cloud peuvent-Ils exploiter la fédération pour réduire la consommation électrique, tout en étant en mesure de fournir un service garantissant les normes de qualité préétablies ? À partir de ces questions, la présente thèse propose une solution de consolidation d’applications pour la fédération de clouds qui garantit le respect des normes de qualité de service. On utilise un système multi-Agents pour négocier la migration des machines virtuelles entre les clouds. En nous basant sur la fédération de clouds, nous avons développé et évalué une approche pour exécuter une énorme application de bioinformatique à coût zéro. En outre, nous avons pu réduire le temps d’exécution de 22,55% par rapport à la meilleure exécution dans un cloud individuel. Cette thèse présente aussi une architecture de cloud baptisée « Excalibur » qui permet l’adaptation automatique des applications standards pour le cloud. Dans l’exécution d’une chaîne de traitements de la génomique, Excalibur a pu parfaitement mettre à l’échelle les applications sur jusqu’à 11 machines virtuelles, ce qui a réduit le temps d’exécution de 63% et le coût de 84% par rapport à la configuration de l’utilisateur. Enfin, cette thèse présente un processus d’ingénierie des lignes de produits (PLE) pour gérer la variabilité de l’infrastructure à la demande du cloud, et une architecture multi-Cloud autonome qui utilise ce processus pour configurer et faire face aux défaillances de manière indépendante. Le processus PLE utilise le modèle étendu de fonction avec des attributs pour décrire les ressources et les sélectionner en fonction des objectifs de l’utilisateur. Les expériences réalisées avec deux fournisseurs de cloud différents montrent qu’en utilisant le modèle proposé, les utilisateurs peuvent exécuter leurs applications dans un environnement de clouds fédérés, sans avoir besoin de connaître les variabilités et contraintes du cloud.

Recent advancements have highlighted that large language models (LLMs), when given a small set of task-specific examples, demonstrate remarkable proficiency, a capability that extends to complex reasoning tasks. In particular, the combination of few-shot learning with the chain-of-thought (CoT) approach has been pivotal in steering models towards more logically consistent conclusions (Wei et al., 2022b). This paper explores the optimization of example selection for designing effective CoT pre-prompts and shows that the choice of the optimization algorithm, typically in favor of comparison-based methods such as evolutionary computation, significantly enhances efficacy and feasibility. Specifically, thanks to a limited exploitative and overfitted optimization, Evolutionary Pre-Prompt Optimization (EPPO) brings an improvement over the naive few-shot approach, exceeding 10 absolute points in exact match scores on benchmark datasets such as GSM8k and MathQA. These gains are consistent across various contexts and are further amplified when integrated with self-consistency (SC).

We propose Partition Trees, a tree-based framework for conditional density estimation over general outcome spaces, supporting both continuous and categorical variables within a unified formulation. Our approach models conditional distributions as piecewise-constant densities on data adaptive partitions and learns trees by directly minimizing conditional negative log-likelihood. This yields a scalable, nonparametric alternative to existing probabilistic trees that does not make parametric assumptions about the target distribution. We further introduce Partition Forests, an ensemble extension obtained by averaging conditional densities. Empirically, we demonstrate improved probabilistic prediction over CART-style trees and competitive or superior performance compared to state-of-the-art probabilistic tree methods and Random Forests, along with robustness to redundant features and heteroscedastic noise.

After Learning Evolutionary Retrofitting (AfterLearnER) consists in applying evolutionary optimization to refine fully trained machine learning models by optimizing a set of carefully chosen parameters or hyperparameters of the model, with respect to some actual, exact, and hence possibly non-differentiable error signal, performed on a subset of the standard validation set. The efficiency of AfterLearnER is demonstrated by tackling non-differentiable signals such as threshold-based criteria in depth sensing, the word error rate in speech resynthesis, the number of kills per life at Doom, computational accuracy or BLEU in code translation, image quality in 3D generative adversarial networks (GANs), and user feedback in image generation via latent diffusion models (LDM). This retrofitting can be done after training, or dynamically at inference time by taking into account the user feedback. The advantages of AfterLearnER are its versatility, the possibility to use non-differentiable feedback, including human evaluations (i.e., no gradient is needed), the limited overfitting supported by a theoretical study, and its anytime behavior. Last but not least, AfterLearnER requires only a small amount of feedback, i.e., a few dozen to a few hundred scalars, compared to the tens of thousands needed in most related published works.

AfterLearnER (After Learning Evolutionary Retrofitting) consists in applying evolutionary optimization to refine fully trained machine learning models by optimizing a set of carefully chosen parameters or hyperpa- rameters of the model, with respect to some actual, exact, and hence possibly non-differentiable error signal, performed on a subset of the standard validation set. The efficiency of AfterLearnER is demonstrated by tackling non-differentiable signals such as threshold-based criteria in depth sensing, the word error rate in speech re-synthesis, the number of kills per life at Doom, computational accuracy or BLEU in code translation, image quality in 3D generative adversarial networks (GANs), and user feedback in image generation via Latent Diffusion Models (LDM). This retrofitting can be done after training, or dynamically at inference time by taking into account the user feedback. The advantages of AfterLearnER are its versatility, the possibility to use non-differentiable feedback, including human evaluations (i.e., no gradient is needed), the limited overfitting supported by a theoretical study, and its anytime behavior. Last but not least, AfterLearnER requires only a small amount of feedback, i.e., a few dozen to a few hundred scalars, compared to the tens of thousands needed in most related published works.

Multi-modal relation extraction (MMRE) is a challenging task that seeks to identify relationships between entities with textual and visual attributes. However, existing methods struggle to handle the complexities posed by multiple entity pairs within a single sentence that share similar contextual information (e.g., identical text and image content). These scenarios amplify the difficulty of distinguishing relationships and hinder accurate extraction. To address these limitations, we propose the variational multi-modal hypergraph attention network (VM-HAN), a novel and robust framework for MMRE. Unlike previous approaches, VM-HAN constructs a multi-modal hypergraph for each sentence-image pair, explicitly modeling high-order intra-/inter-modal correlations among different entity pairs in the same context. This design enables a more detailed and nuanced understanding of entity relationships by capturing intricate cross-modal interactions that are often overlooked. Additionally, we introduce the variational hypergraph attention network (V-HAN). This variational attention mechanism dynamically refines the hypergraph structure, enabling the model to effectively handle the inherent ambiguity and complexity of multi-modal data. Comprehensive experiments on benchmark MMRE datasets demonstrate that VM-HAN achieves state-of-the-art performance, significantly surpassing existing methods in both accuracy and efficiency.

This paper provides a comprehensive review of deep structural causal models (DSCMs), particularly focusing on their ability to answer counterfactual queries using observational data within known causal structures. It delves into the characteristics of DSCMs by analyzing the hypotheses, guarantees, and applications inherent to the underlying deep learning components and structural causal models, fostering a finer understanding of their capabilities and limitations in addressing different counterfactual queries. Furthermore, it highlights the challenges and open questions in the field of deep structural causal modeling. It sets the stages for researchers to identify future work directions and for practitioners to get an overview in order to find out the most appropriate methods for their needs.

This paper introduces Interactive Latent Diffusion Model (IELDM), an encapsulation of a popular text-to-image diffusion model into an Evolutionary framework, allowing the users to steer the design of images toward their goals, alleviating the tedious trial-and-error process that such tools frequently require. The users can not only designate their favourite images, allowing the system to build a surrogate model based on their goals and move in the same directions, but also click on some specific parts of the images to either locally refine the image through dedicated mutation, or recombine images by choosing on each one some regions they like. Experiments validate the benefits of IELDM, especially in a situation where Latent Diffusion Model is challenged by complex input prompts.