Portail HAL des publications du LIRMM
Not a member yet
13279 research outputs found
Sort by
Explainable AI for Embedded Systems Design: A Case Study of Static Redundant NVM Memory Write Prediction
No full textThis paper investigates the application of eXplainable Artificial Intelligence (XAI) in the design of embedded systems using machine learning (ML). As a case study, it addresses the challenging problem of static silent store prediction. This involves identifying redundant memory writes based only on static program features. Eliminating such stores enhances performance and energy efficiency by reducing memory access and bus traffic, especially in the presence of emerging non-volatile memory technologies. To achieve this, we propose a methodology consisting of: 1) the development of relevant ML models for explaining silent store prediction, and 2) the application of XAI to explain these models. We employ two state-of-the-art model-agnostic XAI methods to analyze the causes of silent stores. Through the case study, we evaluate the effectiveness of the methods. We find that these methods provide explanations for silent store predictions, which are consistent with known causes of silent store occurrences from previous studies. Typically, this allows us to confirm the prevalence of silent stores in operations that write the zero constant into memory, or the absence of silent stores in operations involving loop induction variables. This suggests the potential relevance of XAI in analyzing ML models' decision in embedded system design. From the case study, we share some valuable insights and pitfalls we encountered. More generally, this study aims to lay the groundwork for future research in the emerging field of XAI for embedded system design
Peerannot: classification for crowdsourced image datasets with Python
No full textInternational audienceCrowdsourcing is a quick and easy way to collect labels for large datasets, involving many workers. However, workers often disagree with each other. Sources of error can arise from the workers’ skills, but also from the intrinsic difficulty of the task. We present peerannot: a Python library for managing and learning from crowdsourced labels for classification. Our library allows users to aggregate labels from common noise models or train a deep learning-based classifier directly from crowdsourced labels. In addition, we provide an identification module to easily explore the task difficulty of datasets and worker capabilities
A Reinforcement Learning Approach for Resolving Inconsistencies in Qualitative Constraint Networks
No full textInternational audienceIn this paper, we present a reinforcement learning approach for resolving inconsistencies in qualitative constraint networks (QCNs). QCNs are typically used in constraint programming to represent and reason about intuitive spatial or temporal relations like x {isinsideo f ∨ overlaps}y. Naturally, QCNs are not immune to uncertainty, noise, or imperfect data that may be present in information, and thus, more often than not, they are hampered by inconsistencies. We propose a multi-armed bandit approach that defines a well-suited ordering of constraints for finding a maximal satisfiable subset of them. Specifically, our learning approach interacts with a solver, and after each trial a reward is returned to measure the performance of the selected action (constraint addition). The reward function is based on the reduction of the solution space of a consistent reconstruction of the input QCN. Early experimental results obtained by our algorithm suggest that we can do better than the state of the art in terms of both effectiveness, viz., lower number of repairs obtained for an inconsistent QCN, and efficiency, viz., faster runtime
On prime scenarios in qualitative spatial and temporal reasoning
No full textInternational audienceThe concept of prime implicant is a fundamental tool in Boolean algebra, which is used in Boolean circuit design and, recently, in explainable AI. This study investigates an analogous concept in qualitative spatial and temporal reasoning, called prime scenario. Specifically, we define a prime scenario of a qualitative constraint network (QCN) as a minimal set of decisions that can uniquely determine solutions of this QCN. We propose in this paper a collection of algorithms designed to address various problems related to prime scenarios, and also show how certain results can be useful for measuring the robustness of a QCN. In addition, we study the relationship between our notions in this paper and the notion of prime sub-QCN in the literature, and establish theoretical results in the process. Further, we devise a language based on the notion of prime scenario for knowledge compilation. Finally, an experimental evaluation is performed with instances of Allen's Interval Algebra and RCC8 to assess the efficiency of our algorithms and, hence, also the difficulty of the newly introduced problems here
Contribution à la Conception et l'équilibre des Robot Humanoides
No full textHumanoid robots have been the subject of numerous studies for several decades, and their large-scale deployment in industrial contexts has never been closer.In particular, their human-like form makes them perfect candidates for performing low-value tasks in environments that are difficult to access for other mobile platforms.Thus, we see them appearing in factories where contacts with their upper body limbs, often underutilized, are required and have significant consequences on balance.In this thesis, we address the issues of humanoid robotics related to multi-contact balance and the use of the robot's upper limbs.First, the multi-contact balance is studied through a centroidal model that allows for quick calculations. For this, a balance region in which the center of mass must be located is used. Its shape depends on the position of the robot's contacts as well as the tolerated accelerations of the center of mass. To use it online, a method that improves its calculation is proposed. It is then integrated into the control of the humanoid robot HRP-4.Then, several multi-contact balance approaches are compared to the previous method. In particular, an approximation of the equilibrium region of the center of mass by its Chebyshev center and the associated ball is studied.In the second phase, the capabilities of humanoid robots are enhanced by integrating tool changers. The robot will then be able to autonomously change its effectors, whether they are hands, feet, or even the head.A study of commercially available tool changer models is being conducted.Two models are integrated into the humanoid platform HRP-4.We then establish the specifications for a tool changer that is better suited to the constraints of a humanoid robot.Finally, based on these constraints, a new design for an end-effector changer is proposed. This design has variations aimed at every end-effector of the robots, including hands, feet, and head.It takes advantage of the dexterity of the humanoid robot to lock and unlock the effectors without needing to add extra actuators.The result is demonstrated again on the HRP-4 robot.Les robots humanoïdes sont l'objet de nombreuses études depuis plusieurs décennies et leur déploiement à grande échelle dans des contextes industriels n'a jamais été aussi proche.En particulier, leur forme similaire à celle de l'humain en fait des candidats parfaits pour la réalisation de tâches à faible valeur ajoutée dans des environnements difficiles d'accès pour d'autres plateformes mobiles.Ainsi, on les voit apparaître dans les usines où les contacts avec les membres supérieurs, souvent non utilisés, sont requis et ont des conséquences importantes sur l'équilibre.Dans cette thèse, nous abordons les problématiques de la robotique humanoïde liées à l'équilibre multi-contacts et à l'utilisation des membres supérieurs du robot.Tout d'abord l'équilibre multi-contacts est étudié au travers d'un modèle centroidal permettant des calculs rapides.Pour cela, une région d'équilibre dans laquelle doit se trouver le centre de masse est utilisée.Sa forme dépend de la position des contacts du robot ainsi que des accélérations du centre de masse tolérées.Afin de pouvoir l'utiliser en ligne, une méthode améliorant son calcul est proposée.Elle est alors intégrée au contrôle du robot humanoïde HRP-4.Ensuite, plusieurs approches d'équilibre en multi-contacts sont comparées à la méthode précédente.En particulier, une approximation de la région d'équilibre du centre de masse par son centre de Chebyshev et la boule associée est étudiée.Dans un second temps, les capacités des robots humanoïdes sont étendues en leur intégrant des échangeurs d'outils.Le robot pourra alors changer ses effecteurs de manière autonome, que ce soit les mains, les pieds ou encore la tête.Une étude des modèles d'échangeurs disponibles dans le commerce est réalisée.Deux modèles sont intégrés à la plateforme humanoïdes HRP-4.Nous établissons alors le cahier des charges d'un échangeur d'outils plus adapté aux contraintes d'un robot humanoïde.Enfin, à partir de ces contraintes, une nouvelle conception d'échangeur d'outil est proposée.Elle profite de la dextérité du robot humanoïde pour verrouiller et déverrouiller les effecteurs sans nécessiter l'ajout d'actionneur supplémentaires.Le résultat est démontré de nouveau sur le robot HRP-4
A soft skin with self-decoupled three-axis force-sensing taxels
No full textInternational audienceElectronic skins integrating both normal and shear force per taxel have a wide range of applications across diverse fields, e.g., robotics, haptics, and health monitoring. Current multi-axis tactile sensors often present complexities in structure and fabrication or require an extensive calibration process, limiting their widespread applications. In this paper, we report an electronic soft magnetic skin capable of self-decoupling three-axis forces at each taxel. It uses a simple sensor structure with customizable sensitivity and measurement range, reducing the calibration complexity from known quadratic (N 2 ) or cubic (N 3 ) scales down to a linear (3N) scale. The three-axis selfdecoupling property of the sensor is achieved by overlaying two sinusoidally magnetized flexible magnetic films with orthogonal magnetization patterns. Leveraging the self-decoupling feature and its simple structure, we demonstrate that our sensor can facilitate a diverse range of applications, such as measuring the three-dimensional force distribution in artificial knee joints, teaching robots by touch demonstration, and monitoring the interaction forces between knee braces and human skin during various activities
Robust Localization Algorithm for Indoor Robots Based on the Branch-and-Bound Strategy
No full textInternational audienceRobust and accurate localization is crucial for mobile robot navigation in complex indoorenvironments. This paper introduces a robust and integrated robot localization algorithm designedfor such environments. The proposed algorithm, named Branch-and-Bound for Robust Localization(BB-RL), introduces an innovative approach that seamlessly integrates global localization, positiontracking, and resolution of the kidnapped robot problem into a single, comprehensive framework. Theprocess of global localization in BB-RL involves a two-stage matching approach, moving from a broadto a more detailed analysis. This method combines a branch-and-bound algorithm with an iterativenearest point algorithm, allowing for an accurate initial estimation of the robot’s position. For ongoingposition tracking, BB-RL uses a local map-based scan matching technique. To address inaccuracies thataccumulate over time in the local maps, the algorithm creates a pose graph which helps in loop-closureoptimization. Additionally, to make loop-closure detection less computationally intensive, the branch-and-bound algorithm is used to speed up finding loop constraints. A key feature of BB-RL is its FiniteState Machine (FSM)-based relocalization judgment method, which is designed to quickly identify andresolve the kidnapped robot problem. This enhances the reliability of the localization process. BB-RL’sperformance was thoroughly tested in real-world situations using commercially available logisticsrobots. These tests showed that BB-RL is fast, accurate, and robust, making it a practical solution forindoor robot localization
PROSAC as a selection tool for SO-PLS regression: A strategy for multi-block data fusion
No full textInternational audienceBackground: Spectral data from multiple sources can be integrated into multi-block fusion chemometric models, such as sequentially orthogonalized partial-least squares (SO-PLS), to improve the prediction of sample quality features. Pre-processing techniques are often applied to mitigate extraneous variability, unrelated to the response variables. However, the selection of suitable pre-processing methods and identification of informative data blocks becomes increasingly complex and time-consuming when dealing with a large number of blocks. The problem addressed in this work is the efficient pre-processing, selection, and ordering of data blocks for targeted applications in SO-PLS. Results: We introduce the PROSAC-SO-PLS methodology, which employs pre-processing ensembles with response-oriented sequential alternation calibration (PROSAC). This approach identifies the best pre-processed data blocks and their sequential order for specific SO-PLS applications. The method uses a stepwise forward selection strategy, facilitated by the rapid Gram-Schmidt process, to prioritize blocks based on their effectiveness in minimizing prediction error, as indicated by the lowest prediction residuals. To validate the efficacy of our approach, we showcase the outcomes of three empirical near-infrared (NIR) datasets. Comparative analyses were performed against partial-least-squares (PLS) regressions on single-block pre-processed datasets and a methodology relying solely on PROSAC. The PROSAC-SO-PLS approach consistently outperformed these methods, yielding significantly lower prediction errors. This has been evidenced by a reduction in the root-mean-squared error of prediction (RMSEP) ranging from 5 to 25 % across seven out of the eight response variables analyzed. Significance: The PROSAC-SO-PLS methodology offers a versatile and efficient technique for ensemble preprocessing in NIR data modeling. It enables the use of SO-PLS minimizing concerns about pre-processing sequence or block order and effectively manages a large number of data blocks. This innovation significantly streamlines the data pre-processing and model-building processes, enhancing the accuracy and efficiency of chemometric models
Random Schreier Graphs as Expanders
No full textInternational audienceExpander graphs, due to their mixing properties, are useful in many algorithms and combinatorial constructions. One can produce an expander graph with high probability by taking a random graph (e.g., the union of d random bijections for a bipartite graph of degree d). This construction is much simpler than all known explicit constructions of expanders and gives graphs with good mixing properties (small second largest eigenvalue) with high probability. However, from the practical viewpoint, it uses too many random bits, so it is difficult to generate and store these bits for large graphs. The natural idea is to restrict the class of the bijections that we use. For example, if both sides are linear spaces over a finite field , we may consider only linear bijections, making the number of random bits polynomial in k (and not ).In this paper we provide some experimental data that shows that this approach conserves the mixing properties (the second eigenvalue) for several types of graphs (undirected regular and biregular bipartite graphs). We also prove some upper bounds for the second eigenvalue (though they are quite weak compared with the experimental results). Finally, we discuss the possibility to decrease the number of random bits further by using Toeplitz matrices; our experiments show that this change makes the mixing properties only marginally worse while the number of random bits decreases significantly
Kick the Cliques
No full textInternational audienceIn the K_r-Hitting problem, given a graph G and an integer k one has to decide if there exists a set of at most k vertices whose removal destroys all r-cliques of G. In this paper we give an algorithm for K_r-Hitting that runs in subexponential FPT time on graph classes satisfying two simple conditions related to cliques and treewidth. As an application we show that our algorithm solves K_r-Hitting in time - 2^{O_r(k^{(r+1)/(r+2)}log k)} ⋅ n^{O_r(1)} in pseudo-disk graphs and map-graphs; - 2^{O_{t,r}(k^{2/3}log k)} ⋅ n^{O_r(1)} in K_{t,t}-subgraph-free string graphs; and - 2^{O_{H,r}(k^{2/3}log k)} ⋅ n^{O_r(1)} in H-minor-free graphs