HAL-CentraleSupelec
Not a member yet
77624 research outputs found
Sort by
Compressing image encoders via latent distillation
No full textDeep learning models for image compression often face practical limitations in hardware-constrained applications. Although these models achieve high-quality reconstructions, they are typically complex, heavyweight, and require substantial training data and computational resources. We propose a methodology to partially compress these networks by reducing the size of their encoders. Our approach uses a simplified knowledge distillation strategy to approximate the latent space of the original models with less data and shorter training, yielding lightweight encoders from heavyweight ones. We evaluate the resulting lightweight encoders across two different architectures on the image compression task. Experiments show that our method preserves reconstruction quality and statistical fidelity better than training lightweight encoders with the original loss, making it practical for resource-limited environments
Investigations of raster layup effect on mechanical and fracture properties of material extruded ABS samples under biaxial loadings
No full textInternational audienceThe mechanical properties of 3D printed parts are influenced by the parameters used during the manufacturing process. Among them, the raster orientation plays an important role. In this study, biaxial tensile tests coupled with digital image correlation are used to evaluate the effect of raster orientation on the deformation and failure mechanisms of acrylonitrile butadiene styrene samples. It is shown that the local deformation mechanisms are mainly driven by the raster orientation after the crack initiation at sample corners. The kinematic fields evidence that the local deformation mechanisms are related to the mesostructures induced by the fabrication process. A comparative analysis of the different printing orientations (0°/90°, ±45°, 15°/105°, and 30°/120°) reveals that the orthotropic architecture strongly influences the mechanical response and failure mechanisms. This study points out the necessity of improving the filament junctions of polymer parts obtained by fused filament fabrication
Algebraizing Higher-Order Effects
No full textInternational audienceWe present a technique for expressing higher-order effects as algebraic ones. Algebraic effects provide a clean and modular account of computational effects, but they exclude higher-order effects such as local or catch. The standard representation of higher-order effects breaks the separation between syntax and interpretation that algebraic effects rely on. Our proposal, effect algebraization, transforms higher-order effects into combinations of algebraic effects. Each higherorder effect is split into a pair of operations-Open and Close-that together capture the same behavior using only algebraic constructs. The approach is illustrated on reader effects, ask and local. We also sketch a proof of semantics preservation: for every interpretation of the original higher-order effects, there exists a corresponding interpretation of the algebraized ones yielding the same result
Evolutionary Retrofitting
No full textInternational audienceAfterLearnER (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
Visualization experiment and numerical analysis of supercritical CO2 flow inside porous chip: Effect of heat transfer and porous structures
No full textInternational audienc
Quantitative approximation of a Keller–Segel PDE by a branching moderately interacting particle system and suppression of blow-up
No full textThe Keller–Segel PDE is a model for chemotaxis known to exhibit possible finite-time blow-up. Following a seminal work by Tello and Winkler [43], a logistic damping term is added inthis PDE and local well-posedness of mild solutions is proven. When the space dimension is2 or when the damping is strong enough, the solution is global in time. In the second partof this work, a microscopic description of this model is introduced in terms of a system ofstochastic moderately interacting particles. This system features two main characteristics:the interaction between particles happens through a singular (Coulomb-type) kernel which isattractive; and the particles are subject to demographic events, birth and death due to localcompetition with other particles. The latter induces a branching structure of the particlesystem. Then the main result of this work is the convergence of the empirical measure ofthe particle system towards the Keller–Segel PDE with logistic damping, with a rate of orderN − 12(d+1)
Reinforcement Learning-Based Antenna and Time Adaptation Energy-Efficient Strategies in 6G Networks
No full textInternational audienceAmong the strategies for improving the energy efficiency of 5G and beyond networks, sleep-mode control remains one of the most promising techniques for reducing the power consumption of base stations (BSs). The recently standardized 3GPP TR 38.864 BS energy model provides a realistic framework for analyzing multi-level sleep modes and energy-saving mechanisms in next-generation networks. Building upon this model, this paper investigates joint antenna and time adaptation for energy-efficient BS control. The proposed approach allows the BS to autonomously adjust both its temporal activity and antenna usage according to the traffic load. To this end, we design a reinforcement learning (RL) framework in which the BS acts as an intelligent agent that jointly selects the sleepmode level and the number of active antennas. The cost function balances energy consumption and achievable throughput, while penalizing excessive power use and rate shortages. Simulation results based on the 3GPP TR 38.864 power model show that the proposed joint antenna-time adaptation achieves up to 50% power savings compared to time-only adaptation and about 94% compared to antenna-only adaptation, while maintaining satisfactory throughput performance. These results confirm that combining antenna deactivation with time adaptation provides an efficient and scalable solution for energy-aware 6G base-station control
Recouvrements de données dans les protocoles IPv4, IPv6 et TCP : exploration des réassemblages de divers piles réseaux et supervision réseau
No full textIP fragmentation and TCP segmentation enable the splitting of large data packets into smaller ones. These mechanisms permit full or partial overlaps with different data on the overlapping portions. Reassembly policies, i.e., the data chunk preferences that depend on the overlap types, differ across IPv4, IPv6, and TCP implementations. This can lead to vulnerabilities, as Network Intrusion Detection Systems (NIDSes) may interpret the packet differently from the monitored hosts. The main goal of this thesis is to evaluate to what extent NIDSes are vulnerable to overlap-based attacks. We first propose a new way to model the fragment and segment overlaps to ensure complete testing. We instanciate this model in our tool PYROLYSE and test the reassembly policy of various IP and TCP stack types. We discovered that the policies are much more diverse and complex that described by related works. Suricata, Snort and Zeek NIDSes exhibit reassembly inconsistencies, which make them vulnerable to overlap-based attacks. We also found reassembly errors in five stacks, including one CVE.La fragmentation IP et la segmentation TCP permettent de diviser des paquets réseaux trop volumineux en morceaux plus petits. Ce découpage peut donner lieu à du recouvrement, c'est-à-dire que plusieurs morceaux ainsi créés peuvent se chevaucher, de manière complète ou partielle, avec des données non nécessairement identiques. Les politiques de réassemblage, c'est-à-dire le morceau de données préféré en fonction du type de recouvrement, diffèrent selon les implémentations IPv4, IPv6 et TCP. Dès lors, un système de détection d'intrusion réseau (NIDS) qui ne ré-assemble pas les recouvrements de la même manière que l'hôte surveillé est aveugle au flux réellement traité par cet hôte, laissant la place à son contournement. L'objectif principal de cette thèse est d'évaluer dans quelle mesure les NIDS sont vulnérables à des attaques basées sur les recouvrements IPv4, IPv6 et TCP. Nous proposons tout d'abord une nouvelle méthode pour modéliser les recouvrements de fragments et de segments afin de garantir la complétude des tests. Nous instancions ce modèle dans notre outil PYROLYSE et testons les politiques de réassemblage de différents types de piles IP et TCP. Nous avons découvert que les politiques sont beaucoup plus diverses et complexes que décrites dans l'état de l'art et que les NIDS Suricata, Snort et Zeek présentent des incohérences de réassemblage avec ces piles, ce qui les rend vulnérables aux attaques par recouvrement. Nous avons également trouvé des erreurs de réassemblage dans cinq piles, dont une CVE
A Standards-Based Knowledge Graph that Bridges Scientific Workflows, Run-Time Provenance, and Tool Registries
No full textInternational audienceLife science workflows are now prevalent for implementing, executing, and sharing complex data analyses, increasing their scalability and reproducibility. Adhering to the FAIR principles for software further reinforces their reproducibility and the reliability of their results. To maximize their FAIRness, consistent and standardised annotations are critical across several levels: workflows, individual steps, software tools, and input/output data. Such comprehensive metadata make workflows easier to understand, reuse and reproduce, while keeping track of the provenance of their results. However, a unified, queryable knowledge framework that integrates workflows with enriched metadata is lacking. To address this, we developed an integrated workflow knowledge base, that consolidates FAIR metadata from diverse sources and workflow languages into a standardised graph-based representation. It leverages established ontologies and standards (e.g. EDAM, schema.org) to enrich metadata, and link the workflow structure with its execution traces. Our approach provides FAIR-compliant metadata of publicly available pipelines, enabling queries at every granularity level, while accounting for the quality of source data annotation