HAL-INSA Toulouse
Not a member yet
34325 research outputs found
Sort by
Memguard-RW: Improved Real-Time Memory Bandwidth Regulation Within a Hypervisor
No full textInternational audienceMemory bandwidth is a critical factor in the performance of DRAM-based computing architectures, particularly in memory-intensive computations. Modern multi-core processors share critical resources, such as main memory and cache, which impact the predictability of real-time systems due to resource contention. Techniques like memory access regulation, cache partitioning, and static hypervisors aim to mitigate this contention.This paper presents an improvement of a memory control mechanism based on MemGuard, named MemGuard-RW, designed and implemented within a hypervisor. MemGuard-RW uses two performance counters for measuring the memory accesses, one for memory readings and another one for writings, thus decreasing the pessimism on the original memory access budget from MemGuard. Additionally, we extend the MemGuard schedulability analysis considering the two-counter approach. To evaluate the effectiveness of the implementation and analysis, we deployed FreeRTOS as a guest alongside three stress-generating guests, measuring the interference experienced by the FreeRTOS instance and comparing the analysis with two counters with the original analysis with one counter, using modern benchmarks. The results demonstrate that the proposed mechanism successfully regulates memory accesses, showing its potential for enhancing the predictability and performance of real-time systems in multi-core environments. Our proposed analysis with two counters reduces the upper bound of around 20% for tasks having medium and high memory usage
Reconstructing the evolution of cerebral reorganisation in the hominin fossil record
No full textInternational audienc
Numerical modelling of thermoplastic CO2 assisted extrusion foaming: Exploring the process-microstructure relationship
No full textInternational audienceThis study presents a comprehensive numerical modeling approach for thermoplastic foams processing, produced via supercritical CO₂-assisted extrusion foaming. The model focuses on the nucleation and growth of bubbles within the polymer matrix, employing the influence volume approach to simulate these dynamics. Key phenomena such as gas diffusion, nucleation rate, bubble growth, and the evolution of influence and non-influence volumes are captured to predict foam microstructure and cellular density. Results highlight the model’s capability to predict equilibrium bubble sizes ranging from micrometers to hundreds of micrometers, offering insights into foam behavior and microstructure control. The final size distribution will be used to create on Paraview® the microstructure that will be compared to experimental observations in the future
Open-Hole tensile tests on poplar veneer laminates and plywood: Applicability of the point stress criterion
No full textInternational audiencehis study explores the open-hole tensile strength of two types of poplar veneer laminates: a quasi-isotropic laminate [90°/45°/0°/–45°]s and a plywood laminate [90°/0°/90°/0°1/2]s. The effects of hole diameter and specimen geometry on tensile strength were examined, and the Point Stress Criterion (PSC) was used to evaluate material behaviours. The results show that the stress concentration values (d0) for the veneer laminates align closely with those of fibre-reinforced polymer composites, demonstrating the applicability of the PSC to wood-based laminates.While the d0 values are similar to those observed in synthetic composites, the failure mode differs significantly. Unlike most synthetic composites, no delamination occurs near the hole or along the edges of the studied laminates. Instead, the specimens exhibit a brittle fracture mode, characterised by sudden failure without ply separation.These findings provide insights into the influence of hole effects on laminate design and suggest that the PSC can be used to optimise the performance of veneer laminates in engineering applications
Méthodes de résolution de processus de décision markoviens multi-objectifs à horizon fini
No full textThis dissertation develops methods for the optimization of a discrete-time Markov decision process (MDP) characterized by dollar k geq 2 dollar objective functions. At the beginning of each time period, the MDP occupies a state in which a number of alternative actions are available. The choice of an action generates a dollar k dollar -dimensional reward and determines the probability distribution of the next state. A policy specifies the action to be taken in each state as a function of the period. A policy that achieves a Pareto efficient expected total reward vector over the process's lifetime is termed efficient. A uniformly efficient policy is a policy efficient for all initial states. This study deals strictly with the case of a finite time horizon. It is divided into three main chapters. In the first chapter, we disprove the standard theorem on the vector extension of the optimality equations of a finite-horizon single-objective MDP. By dint of a counterexample, we show that solution of the extended equations does not always yield the efficient value sets stipulated by the theorem. Analysis of the counterexample leads us to a condition sufficient for the theorem to hold in its original form. The condition is shown to hold when one or two time periods are left, as well as when the dynamics of the model are deterministic. More generally, we demonstrate that the actual solutions to the equations are the sets of all efficient values achieved in a larger policy class than that allowed by the theorem. This class includes Markov and non-Markov policies. We describe a dynamic programming algorithm for constructing uniformly efficient Markov policies with respect to the new policy space. The second chapter focuses on how to characterize and compute uniformly efficient policies when we only allow consideration of Markov deterministic policies. To this end, the concept of an F-efficient policy is introduced. We show F-efficiency to be a necessary condition for uniform efficiency that can be characterized using functional equations. The equations yield a dynamic programming algorithm that finds all F-efficient policies. From an alternate representation of the uniformly efficient policy set within the F-efficient policy set, we deduce a procedure for enumerating the former set. Numerical experiments with this procedure suggest definite computational advantages over full search-based policy optimization. The third chapter proposes a multi-objective linear programming formulation of a model instance in which rewards and transition probabilities can vary with time. We allow consideration of all Markov randomized policies, but in order to exclude policies of a certain anomalous type, we introduce the concept of a regular policy. We establish a one-to-one correspondence between the efficient solutions of the multi-objective linear program and the efficient regular policies. Thanks to this equivalence, we are able to characterize, and demonstrate the existence of, efficient deterministic policies. The characterization is then employed in the development of a simplex method-like algorithm that detects all efficient deterministic policies after a finite number of iterations. To our knowledge, this is the only algorithm in print designed for accomplishing this end. We report the results of testing the algorithm in a bi-objective engineering application. The methods presented in this study enable and inform the resolution of decision-making problems that can be treated as finite-horizon MDPs but cannot realistically be represented by a single objective function.Cette thèse développe des méthodes pour l'optimisation d'un processus de décision markovien (MDP) à temps discret, caractérisé par dollar k geq 2 dollar fonctions objectif. Au début de chaque période, le MDP se trouve dans un état où plusieurs actions alternatives sont disponibles. Le choix d'une action génère une récompense de taille dollar k dollar et détermine la distribution de probabilité de l'état suivant. Une politique spécifie l'action à entreprendre dans chaque état. Une politique est dite efficiente si elle réalise un vecteur de récompense total espéré Pareto efficient sur l'horizon du processus. Une politique uniformément efficiente est une politique qui est efficiente quel que soit l'état initial. Cette étude se limite au cas d'un horizon temporel fini. Elle est divisée en trois chapitres. Dans le premier chapitre, nous réfutons le théorème standard portant sur l'extension vectorielle des équations d'optimalité d'un MDP mono-objectif à horizon fini. Au moyen d'un contre-exemple, nous montrons que la solution des équations étendues ne fournit pas toujours les ensembles de valeurs efficientes stipulés par le théorème. L'analyse du contre-exemple nous conduit à une condition suffisante pour que le théorème soit valable dans sa forme originelle. Cette condition est démontrée vraie lorsqu'il ne reste qu'une ou deux périodes, ainsi que lorsque la dynamique du modèle est déterministe. Plus généralement, nous démontrons que les vraies solutions des équations correspondent aux ensembles des valeurs efficientes atteintes dans une classe plus large de politiques que celle autorisée par le théorème. Cette classe inclut des politiques markoviennes et non markoviennes. Nous décrivons un algorithme de programmation dynamique permettant de construire des politiques uniformément efficientes par rapport au nouvel espace de politiques. Le deuxième chapitre s'attache à caractériser et à calculer les politiques uniformément efficientes lorsqu'on ne considère que les politiques markoviennes déterministes. À cette fin, le concept de politique F-efficiente est introduit. Nous montrons que la F-efficience est une condition nécessaire à l'efficience uniforme, et qu'elle peut être caractérisée à l'aide d'équations d'optimalité. Ces équations donnent lieu à un algorithme de programmation dynamique qui identifie toutes les politiques F-efficientes. À partir d'une certaine représentation de l'ensemble des politiques uniformément efficientes au sein de l'ensemble F-efficient, nous déduisons une procédure permettant d'énumérer les politiques uniformément efficientes. Des expériences numériques menées avec cette procédure suggèrent des avantages computationnels par rapport à une recherche exhaustive. Le troisième chapitre propose une formulation en programmation linéaire multi-objectif d'un cas de figure dans lequel les récompenses et les probabilités de transition peuvent varier dans le temps. Nous établissons une bijection entre les solutions efficientes du programme linéaire et les politiques efficientes. Grâce à cette équivalence, nous sommes en mesure de caractériser et de démontrer l'existence de politiques déterministes efficientes. Cette caractérisation est ensuite utilisée dans le développement d'un algorithme de type simplexe permettant de détecter toutes les politiques déterministes efficientes en un nombre fini d'itérations. À notre connaissance, il s'agit du seul algorithme dans la littérature qui est conçu spécialement pour cet objectif. Nous présentons les résultats de son application dans un problème d'ingénierie bicritère. Les méthodes présentées dans cette thèse facilitent et éclairent la résolution de problèmes de prise de décision pouvant être modélisés comme des MDP à horizon fini mais ne pouvant être représentés de manière réaliste par une unique fonction objectif
Paul de Faget de Casteljau, Un pionnier de la CAGD
No full textInternational audiencewe were saddened by the loss of Paul de Faget de Casteljau, one of the most prominent founding members of the research area Computer Aided Geometric Design (CAGD) and pioneer of modern applied geometry in industry. With this special issue we take the opportunity to celebrate his contributions and to reflect on the impact of his research to the field. After a careful review process, we decided to accept 17 out of the 28 submissions, including 2 articles about Paul de Casteljau's work and life, 5 surveys summarizing the numerous research directions inspired by his ideas, and 10 research articles which showcase that these ideas are still alive and keep stimulating the CAGD community.</div
Image synthesis using a physics-informed diffusion model for crack detection with laser flying spot thermography
No full textInternational audienceIn non-destructive testing for metallic materials, ‘Flying-spot’ thermography is a method for detecting cracks based on the infrared scanning of samples by a local laser heat source. However, identifying a crack from other surface artefacts such as air ducts or non-planar shapes on the material surface can be challenging. That is why a recent series of works has proposed the use of deep learning techniques, which can exploit contextual information but require a significant amount of data. Some of these works use a diffusion model to increase the amount of data available to train the defect detection model. However, the diffusion model only imitates the training images it was fed and therefore, there is no physical meaning enforced to the images it generates. This can cause some synthetic images to exhibit temperature fields that do not comply with the diffusion law, or even contain obvious inconsistencies. In this paper, we propose the implementation of a physics-informed diffusion model to generate synthetic thermal images with a real physical meaning. In the literature, the integration of physical information is achieved using a partial differential equation (PDE). In contrast, we propose to directly incorporate the solution of the equation for a semi-infinite homogeneous solid (i.e. a defect-free material). To evaluate the benefits of such a model, several metrics are measured such as the Fréchet Inception Distance, the physical error and the speed of convergence, but also the performance on a downstream anomaly detection task
Approximation par réseaux de neurones et construction de réseaux de neurones récurrents quantifiés
No full textIn this thesis, we address two important issues in deep learning: approximation and compression. Regarding approximation, we study the expressivity of deep neural networks, which is the focus of Chapter 3 of this thesis. Regarding compression, we study the quantization of orthogonal recurrent neural networks and we introduce two models, QORNN and HadamRNN. This is developed in Chapters 4 and 5. Approximation: We study the fundamental limits to the expressive power of neural networks. Given two sets F, G of real-valued functions, we first prove a general lower bound on how well functions in F can be approximated in Lp(mu) norm by functions in G, for any p >= 1 and any probability measure mu. The lower bound depends on the packing number of F, the range of F, and the fat-shattering dimension of G. We then instantiate this bound to the case where G corresponds to a piecewise-polynomial feed-forward neural network, and describe in details the application to two sets F: Hölder balls and multivariate monotonic functions. Beside matching (known or new) upper bounds up to log factors, our lower bounds shed some light on the similarities or differences between approximation in Lp norm or in sup norm, solving an open question by DeVore et al. Our proof strategy differs from the sup norm case and uses a key probability result of Mendelson. QORNN: We explore the quantization of the weight matrices in ORNNs, leading to Quantized approximately Orthogonal RNNs (QORNNs). The construction of such networks remained an open problem, acknowledged for its inherent instability. We propose and investigate two strategies to learn QORNN by combining quantization-aware training (QAT) and orthogonal projections. We also study post-training quantization of the activations for pure integer computation of the recurrent loop. The most efficient models achieve results similar to state-of-the-art full-precision ORNN, LSTM and FastRNN on a variety of standard benchmarks, even with 4-bits quantization. HadamRNN: Binary and sparse ternary weights in neural networks enable faster computations and lighter representations, facilitating their use on edge devices with limited computational power. Meanwhile, vanilla RNNs are highly sensitive to changes in their recurrent weights, making the binarization and ternarization of these weights inherently challenging. To date, no method has successfully achieved binarization or ternarization of vanilla RNN weights. We present a new approach leveraging the properties of Hadamard matrices to parameterize a subset of binary and sparse ternary orthogonal matrices. This method enables the training of orthogonal RNNs (ORNNs) with binary and sparse ternary recurrent weights, effectively creating a specific class of binary and sparse ternary vanilla RNNs. The resulting ORNNs, named HadamRNN and Block-HadamRNN, are evaluated on various benchmarks, including the copy task, permuted and sequential MNIST tasks, the IMDB dataset, two GLUE benchmarks, and two IoT benchmarks. Despite binarization or sparse ternarization, these RNNs maintain performance levels comparable to state-of-the-art full-precision models, highlighting the effectiveness of our approach. Notably, our approach is the first solution with binary recurrent weights capable of tackling the copy task over 1000 timesteps.Dans cette thèse, nous abordons deux problématiques importantes en apprentissage profond : l'approximation et la compression. En approximation, nous étudions l'expressivité des réseaux de neurones profonds. Cela fait l'objet du chapitre 3 de cette thèse. En compression, nous étudions la quantification, jusqu'à la binarisation et ternarisation, des réseaux de neurones récurrents orthogonaux, et nous introduisons deux modèles, QORNN et HadamRNN. Cela fait l'objet des chapitres 4 et 5. Approximation : nous étudions les limites fondamentales du pouvoir expressif des réseaux de neurones. Étant donnés deux ensembles F et G de fonctions à valeurs réelles, nous démontrons d'abord une borne inférieure générale sur l'erreur d'approximation des fonctions de F par des fonctions de G en norme Lp(mu), pour tout p >= 1 et toute mesure de probabilité mu. Cette borne inférieure dépend du packing-number de F, de l'amplitude de F et de la fat-shattering dimension de G. Nous appliquons ensuite cette borne au cas où G correspond à un réseau de neurones MLP avec activation polynômiale par morceaux, et nous décrivons en détail son application à deux ensembles F : les boules de Hölder et les fonctions monotones multivariées. En plus d'établir des bornes inférieures qui correspondent aux bornes supérieures (connues ou nouvelles) à un facteur logarithmique près, nos résultats éclairent les similitudes et différences entre l'approximation en norme Lp et en norme sup, répondant ainsi à une question ouverte posée par DeVore et al. Notre stratégie de preuve diffère du cas de la norme sup et s'appuie sur un résultat clé en probabilité de Mendelson. QORNN : nous explorons la quantification des matrices de poids dans les ORNNs, conduisant à des réseaux de neurones récurrents approximativement orthogonaux quantifiés (Quantized Approximately Orthogonal RNNs, QORNNs). La construction de tels réseaux restait un problème ouvert, reconnu pour son instabilité inhérente. Nous proposons et analysons deux stratégies pour apprendre des QORNNs en combinant la quantification pendant l'entrainement (Quantization-Aware Training, QAT) et les projections orthogonales. Nous étudions également la quantification post-entraînement (PTQ) des activations afin de permettre un calcul purement entier de la boucle récurrente. Les modèles les plus efficaces obtiennent des résultats similaires aux ORNNs full-precision (non-quantifiés), ainsi qu'aux LSTM et FastRNN de l'état de l'art sur une variété de benchmarks standards, même avec une quantification en 4 bits. HadamRNN : les RNNs classiques sont extrêmement sensibles aux changements de leurs poids récurrents, rendant la binarisation et la ternarisation de ces poids difficiles. À ce jour, aucune méthode ne permet de binariser ou ternariser les poids récurrents des RNNs classiques. Nous présentons une nouvelle approche, exploitant les propriétés des matrices de Hadamard, pour paramétriser un sous-ensemble de matrices orthogonales binaires et ternaires parcimonieuses. Cette méthode permet l'entraînement de réseaux de neurones récurrents orthogonaux (ORNNs) avec des poids récurrents binaires et ternaires parcimonieux, créant ainsi une classe spécifique de RNNs classiques binaires et ternaires parcimonieux. Les ORNNs obtenus, nommés HadamRNN et Block-HadamRNN, sont évalués sur divers benchmarks, notamment la Copy task, Sequential MNIST (permutée et non permutée), IMDB dataset, deux benchmarks GLUE ainsi que deux benchmarks en IoT. Malgré la binarisation ou la ternarisation parcimonieuse, ces RNNs maintiennent des performances comparables aux modèles full-precision de l'état de l'art, soulignant l'efficacité de notre approche. Notamment, notre approche est la première solution avec des poids récurrents binaires capable de résoudre la Copy task sur plus de 1000 pas de temps
Increasing the throughput of Direct-to-Satellite Narrowband IoT networks
No full textInternational audienceThe design of a Direct-to-Satellite (DtS) network infrastructure based on the availability of Low Earth Orbit (LEO) satellite constellations has recently emerged as a key enabler of global Internet of Things (IoT) connectivity. In this context, Narrowband IoT (NB-IoT) communications can be leveraged as a standards-based DtS solution that can grant reliable data collection in areas not reachable by traditional terrestrial IoT networks. However, the large coverage area of satellites and the limited connection time increase the probability of collisions during the NB-IoT Random Access (RA) procedure. This is especially true when many concurrent User Equipments (UEs) attempt to access the network simultaneously during the RA procedure. In this sense, the goal of this contribution is to show that the first message exchange between UEs and NB-IoT equipped LEO satellites (i.e., an uplink Msg1 frame followed by a downlink Msg2 frame) can be exploited by UEs to decide whether to try transmitting data into an uplink Msg3 frame or not. Herein, 2 early collision detection methods able to improve the RA success rate in satellite NB-IoT systems are presented. In detail, for both methods UEs estimate their relative position to the satellite and apply a corresponding time shift when transmitting Msg1. By comparing the Time Advance (TA) value received in Msg2 with their expected TA, UEs can determine whether to proceed with Msg3. The Closest First Method (CFM) requires no changes to the satellite and only uses the TA value of the first received Msg1. Instead, the Non-collided First Method (NFM) assumes the satellite can identify individual Msg1 transmissions and responds only to non-collided ones. Some preliminary simulation results show that both methods increase the number of successful accesses compared to the standard approach, even in the presence of position estimation errors