Avec le développement de la robotique cognitive, le besoin d’outils avancés pour représenter, manipuler, raisonner sur les connaissances acquises par un robot a clairement été mis en avant. Mais stocker et manipuler des connaissances requiert tout d’abord d’éclaircir ce que l’on nomme connaissance pour un robot, et comment celle-ci peut-elle être représentée de manière intelligible pour une machine. Ce travail s’efforce dans un premier temps d’identifier de manière systématique les besoins en terme de représentation de connaissance des applications robotiques modernes, dans le contexte spécifique de la robotique de service et des interactions homme-robot. Nous proposons une typologie originale des caractéristiques souhaitables des systèmes de représentation des connaissances, appuyée sur un état de l’art détaillé des outils existants dans notre communauté. Dans un second temps, nous présentons en profondeur ORO, une instanciation particulière d’un système de représentation et manipulation des connaissances, conçu et implémenté durant la préparation de cette thèse. Nous détaillons le fonctionnement interne du système, ainsi que son intégration dans plusieurs architectures robotiques complètes. Un éclairage particulier est donné sur la modélisation de la prise de perspective dans le contexte de l’interaction, et de son interprétation en terme de théorie de l’esprit. La troisième partie de l’étude porte sur une application importante des systèmes de représentation des connaissances dans ce contexte de l’interaction homme-robot : le traitement du dialogue situé. Notre approche et les algorithmes qui amènent à l’ancrage interactif de la communication verbale non contrainte sont présentés, suivis de plusieurs expériences menées au Laboratoire d’Analyse et d’Architecture des Systèmes au CNRS à Toulouse, et au groupe Intelligent Autonomous System de l’université technique de Munich. Nous concluons cette thèse sur un certain nombre de considérations sur la viabilité et l’importance d’une gestion explicite des connaissances des agents, ainsi que par une réflexion sur les éléments encore manquant pour réaliser le programme d’une robotique “de niveau humain” / With the rise of the so-called cognitive robotics, the need of advanced tools to store, manipulate, reason about the knowledge acquired by the robot has been made clear. But storing and manipulating knowledge requires first to understand what the knowledge itself means to the robot and how to represent it in a machine-processable way. This work strives first at providing a systematic study of the knowledge requirements of modern robotic applications in the context of service robotics and human-robot interaction. What are the expressiveness requirement for a robot? what are its needs in term of reasoning techniques? what are the requirement on the robot's knowledge processing structure induced by other cognitive functions like perception or decision making? We propose a novel typology of desirable features for knowledge representation systems supported by an extensive review of existing tools in our community. In a second part, the thesis presents in depth a particular instantiation of a knowledge representation and manipulation system called ORO, that has been designed and implemented during the preparation of the thesis. We elaborate on the inner working of this system, as well as its integration into several complete robot control stacks. A particular focus is given to the modelling of agent-dependent symbolic perspectives and their relations to theories of mind. The third part of the study is focused on the presentation of one important application of knowledge representation systems in the human-robot interaction context: situated dialogue. Our approach and associated algorithms leading to the interactive grounding of unconstrained verbal communication are presented, followed by several experiments that have taken place both at the Laboratoire d'Analyse et d'Architecture des Systèmes at CNRS, Toulouse and at the Intelligent Autonomous System group at Munich Technical University. The thesis concludes on considerations regarding the viability and importance of an explicit management of the agent's knowledge, along with a reflection on the missing bricks in our research community on the way towards "human level robots"

Beschreibungslogiken (Description logics, DLs) sind eine Klasse von Wissensrepraesentationsformalismen mit wohldefinierter, logik-basierter Semantik und entscheidbaren Schlussfolgerungsproblemen, wie z.B. dem Erfuellbarkeitsproblem. Zwei wichtige Entscheidungsverfahren fuer das Erfuellbarkeitsproblem von DL-Ausdruecken sind Tableau- und Automaten-basierte Algorithmen. Diese haben aufgrund ihrer unterschiedlichen Arbeitsweise komplementaere Eigenschaften: Tableau-Algorithmen eignen sich fuer Implementierungen und fuer den Nachweis von PSPACE- und NEXPTIME-Resultaten, waehrend Automaten sich besonders fuer EXPTIME-Resultate anbieten. Zudem ermoeglichen sie eine vom Standpunkt der Theorie aus elegantere Handhabung von unendlichen Strukturen, eignen sich aber wesentlich schlechter fuer eine Implementierung. Ziel der Dissertation ist es, die Gruende fuer diese Unterschiede zu analysieren und Moeglichkeiten aufzuzeigen, wie Eigenschaften von einem Ansatz auf den anderen uebertragen werden koennen, um so die positiven Eigenschaften von beiden Ansaetzen miteinander zu verbinden. Unter Anderem werden Methoden entwickelt, mit Hilfe von Automaten PSPACE-Resultate zu zeigen, und von einem Tableau-Algorithmus automatisch ein EXPTIME-Resultat abzuleiten. / Description Logics (DLs) are a family of knowledge representation languages with well-defined logic-based semantics and decidable inference problems, e.g. satisfiability. Two of the most widely used decision procedures for the satisfiability problem are tableau- and automata-based algorithms. Due to their different operation, these two classes have complementary properties: tableau algorithms are well-suited for implementation and for showing PSPACE and NEXPTIME complexity results, whereas automata algorithms are particularly useful for showing EXPTIME results. Additionally, they allow for an elegant handling of infinite structures, but they are not suited for implementation. The aim of this thesis is to analyse the reasons for these differences and to find ways of transferring properties between the two approaches in order to reconcile the positive properties of both. For this purpose, we develop methods that enable us to show PSPACE results with the help of automata and to automatically derive an EXPTIME result from a tableau algorithm.

<p> The dramatic improvement in information and communication technology (ICT) has made an evolution in learning management systems (LMS). The rapid growth in LMSs has caused users to demand more advanced, automated, and intelligent services. CourseNetworking is a next-generation LMS adopting machine learning to add personalization, gamification, and more dynamics to the system. This work tries to come up with two recommender systems that can help improve CourseNetworking services. The first one is a social recommender system helping CourseNetworking to track user interests and give more relevant recommendations. Recently, graph neural network (GNN) techniques have been employed in social recommender systems due to their high success in graph representation learning, including social network graphs. Despite the rapid advances in recommender systems performance, dealing with the dynamic property of the social network data is one of the key challenges that is remained to be addressed. In this research, a novel method is presented that provides social recommendations by incorporating the dynamic property of social network data in a heterogeneous graph by supplementing the graph with time span nodes that are used to define users long-term and short-term preferences over time. The second service that is proposed to add to Rumi services is a hashtag recommendation system that can help users label their posts quickly resulting in improved searchability of content. In recent years, several hashtag recommendation methods are proposed and developed to speed up processing of the texts and quickly find out the critical phrases. The methods use different approaches and techniques to obtain critical information from a large amount of data. This work investigates the efficiency of unsupervised keyword extraction methods for hashtag recommendation and recommends the one with the best performance to use in a hashtag recommender system. </p>

<p dir="ltr">We present a framework for Trustworthy Artificial Intelligence (TAI) that dynamically assesses trust and scrutinizes past decision-making, aiming to identify both individual and community behavior. The modeling of behavior incorporates proposed concepts, namely trust pressure and trust sensitivity, laying the foundation for predicting future decision-making regarding community behavior, consensus level, and decision-making duration. Our framework involves the development and mathematical modeling of trust pressure and trust sensitivity, drawing on social validation theory within the context of environmental decision-making. To substantiate our approach, we conduct experiments encompassing (i) dynamic trust sensitivity to reveal the impact of learning actors between decision-making, (ii) multi-level trust measurements to capture disruptive ratings, and (iii) different distributions of trust sensitivity to emphasize the significance of individual progress as well as overall progress.</p><p dir="ltr">Additionally, we introduce TAI metrics, trustworthy acceptance, and trustworthy fairness, designed to evaluate the acceptance of decisions proposed by AI or humans and the fairness of such proposed decisions. The dynamic trust management within the framework allows these TAI metrics to discern support for decisions among individuals with varying levels of trust. We propose both the metrics and their measurement methodology as contributions to the standardization of trustworthy AI.</p><p dir="ltr">Furthermore, our trustability metric incorporates reliability, resilience, and trust to evaluate systems with multiple components. We illustrate experiments showcasing the effects of different trust declines on the overall trustability of the system. Notably, we depict the trade-off between trustability and cost, resulting in net utility, which facilitates decision-making in systems and cloud security. This represents a pivotal step toward an artificial control model involving multiple agents engaged in negotiation.</p><p dir="ltr">Lastly, the dynamic management of trust and trustworthy acceptance, particularly in varying criteria, serves as a foundation for causal AI by providing inference methods. We outline a mechanism and present an experiment on human-driven causal inference, where participant discussions act as interventions, enabling counterfactual evaluations once actor and community behavior are modeled.</p>

In this thesis, we study advanced reasoning about dynamical systems in a logical framework -- the situation calculus. In particular, we consider promoting the efficiency of reasoning about action in the situation calculus from three different aspects. First, we propose a modified situation calculus based on the two-variable predicate logic with counting quantifiers. We show that solving the projection and executability problems via regression in such language are decidable. We prove that generally these two problems are co-NExpTime-complete in the modified language. We also consider restricting the format of regressable formulas and basic action theories (BATs) further to gain better computational complexity for reasoning about action via regression. We mention possible applications to formalization of Semantic Web services. Then, we propose a hierarchical representation of actions based on the situation calculus to facilitate development, maintenance and elaboration of very large taxonomies of actions. We show that our axioms can be more succinct, while still using an extended regression operator to solve the projection problem. Moreover, such representation has significant computational advantages. For taxonomies of actions that can be represented as finitely branching trees, the regression operator can sometimes work exponentially faster with our theories than it works with the BATs current situation calculus. We also propose a general guideline on how a taxonomy of actions can be constructed from the given set of effect axioms. Finally, we extend the current situation calculus with the order-sorted logic. In the new formalism, we add sort theories to the usual initial theories to describe taxonomies of objects. We then investigate what is the well-sortness for BATs under such framework. We consider extending the current regression operator with well-sortness checking and unification techniques. With the modified regression, we gain computational efficiency by terminating the regression earlier when reasoning tasks are ill-sorted and by reducing the search spaces for well-sorted objects. We also study that the connection between the order-sorted situation calculus and the current situation calculus.

In this thesis, we study advanced reasoning about dynamical systems in a logical framework -- the situation calculus. In particular, we consider promoting the efficiency of reasoning about action in the situation calculus from three different aspects. First, we propose a modified situation calculus based on the two-variable predicate logic with counting quantifiers. We show that solving the projection and executability problems via regression in such language are decidable. We prove that generally these two problems are co-NExpTime-complete in the modified language. We also consider restricting the format of regressable formulas and basic action theories (BATs) further to gain better computational complexity for reasoning about action via regression. We mention possible applications to formalization of Semantic Web services. Then, we propose a hierarchical representation of actions based on the situation calculus to facilitate development, maintenance and elaboration of very large taxonomies of actions. We show that our axioms can be more succinct, while still using an extended regression operator to solve the projection problem. Moreover, such representation has significant computational advantages. For taxonomies of actions that can be represented as finitely branching trees, the regression operator can sometimes work exponentially faster with our theories than it works with the BATs current situation calculus. We also propose a general guideline on how a taxonomy of actions can be constructed from the given set of effect axioms. Finally, we extend the current situation calculus with the order-sorted logic. In the new formalism, we add sort theories to the usual initial theories to describe taxonomies of objects. We then investigate what is the well-sortness for BATs under such framework. We consider extending the current regression operator with well-sortness checking and unification techniques. With the modified regression, we gain computational efficiency by terminating the regression earlier when reasoning tasks are ill-sorted and by reducing the search spaces for well-sorted objects. We also study that the connection between the order-sorted situation calculus and the current situation calculus.

<p>Over the past decade, distributed representation learning has emerged as a popular alternative to conventional centralized machine learning training. The increasing interest in distributed representation learning, specifically federated learning, can be attributed to its fundamental property that promotes data privacy and communication savings. While conventional ML encourages aggregating data at a central location (e.g., data centers), distributed representation learning advocates keeping data at the source and instead transmitting model parameters across the network. However, since the advent of deep learning, model sizes have become increasingly large often comprising million-billions of parameters, which leads to the problem of communication latency in the learning process. In this thesis, we propose to tackle the problem of communication latency in two different ways: (i) learning private representation of data to enable its sharing, and (ii) reducing the communication latency by minimizing the corresponding long-range communication requirements.</p> <p><br></p> <p>To tackle the former goal, we first start by studying the problem of learning representations that are private yet informative, i.e., providing information about intended ''ally'' targets while hiding sensitive ''adversary'' attributes. We propose Exclusion-Inclusion Generative Adversarial Network (EIGAN), a generalized private representation learning (PRL) architecture that accounts for multiple ally and adversary attributes, unlike existing PRL solutions. We then address the practical constraints of the distributed datasets by developing Distributed EIGAN (D-EIGAN), the first distributed PRL method that learns a private representation at each node without transmitting the source data. We theoretically analyze the behavior of adversaries under the optimal EIGAN and D-EIGAN encoders and the impact of dependencies among ally and adversary tasks on the optimization objective. Our experiments on various datasets demonstrate the advantages of EIGAN in terms of performance, robustness, and scalability. In particular, EIGAN outperforms the previous state-of-the-art by a significant accuracy margin (47% improvement), and D-EIGAN's performance is consistently on par with EIGAN under different network settings.</p> <p><br></p> <p>We next tackle the latter objective - reducing the communication latency - and propose two timescale hybrid federated learning (TT-HF), a semi-decentralized learning architecture that combines the conventional device-to-server communication paradigm for federated learning with device-to-device (D2D) communications for model training. In TT-HF, during each global aggregation interval, devices (i) perform multiple stochastic gradient descent iterations on their individual datasets, and (ii) aperiodically engage in consensus procedure of their model parameters through cooperative, distributed D2D communications within local clusters. With a new general definition of gradient diversity, we formally study the convergence behavior of TT-HF, resulting in new convergence bounds for distributed ML. We leverage our convergence bounds to develop an adaptive control algorithm that tunes the step size, D2D communication rounds, and global aggregation period of TT-HF over time to target a sublinear convergence rate of O(1/t) while minimizing network resource utilization. Our subsequent experiments demonstrate that TT-HF significantly outperforms the current art in federated learning in terms of model accuracy and/or network energy consumption in different scenarios where local device datasets exhibit statistical heterogeneity. Finally, our numerical evaluations demonstrate robustness against outages caused by fading channels, as well favorable performance with non-convex loss functions.</p>

<p>Detecting crop rows from video frames in real time is a fundamental challenge in the field of precision agriculture. Deep learning based semantic segmentation method, namely U-net, although successful in many tasks related to precision agriculture, performs poorly for solving this task. The reasons include paucity of large scale labeled datasets in this domain, diversity in crops, and the diversity of appearance of the same crops at various stages of their growth. In this work, we discuss the development of a practical real-life crop row</p> <p>detection system in collaboration with an agricultural sprayer company. Our proposed method takes the output of semantic segmentation using U-net, and then apply a clustering based probabilistic temporal calibration which can adapt to different fields and crops without the need for retraining the network. Experimental results validate that our method can be used for both refining the results of the U-net to reduce errors and also for frame interpolation of the input video stream. Upon the availability of more labeled data, we switched our approach from a semi-supervised model to a fully supervised end-to-end crop row detection model using a Feature Pyramid Network or FPN. Central to the FPN is a pyramid pooling module that extracts features from the input image at multiple resolutions. This results in the network’s ability to use both local and global features in classifying pixels to be crop rows. After training the FPN on the labeled dataset, our method obtained a mean IoU or Jaccard Index score of over 70% as reported on the test set. We trained our method on only a subset of the corn dataset and tested its performance on multiple variations of weed pressure and crop growth stages to verify that the performance does translate over the variations and is consistent across the entire dataset.</p>

<p dir="ltr">Data is collected everywhere in our world; however, it often is noisy and incomplete. Different sources of data may have different characteristics, quality levels, or come from dynamic and diverse environments. This poses challenges for both humans who want to gain insights from data and machines which are learning patterns from data. How can we leverage the diversity of data across domains to enhance our understanding and decision-making? In this thesis, we address this question by proposing novel methods and applications that use multiple domains as more holistic sources of information for both human and machine learning tasks. For example, to help human operators understand environmental dynamics, we show the detection and localization of distribution shifts to problematic features, as well as how interpretable distributional mappings can be used to explain the differences between shifted distributions. For robustifying machine learning, we propose a causal-inspired method to find latent factors that are robust to environmental changes and can be used for counterfactual generation or domain-independent training; we propose a domain generalization framework that allows for fast and scalable models that are robust to distribution shift; and we introduce a new dataset based on human matches in StarCraft II that exhibits complex and shifting multi-agent behaviors. We showcase our methods across various domains such as healthcare, natural language processing (NLP), computer vision (CV), etc. to demonstrate that learning from data across domains can lead to more faithful representations of data and its generating environments for both humans and machines.</p>

<p dir="ltr">Explainable AI (XAI) is a class of post-hoc analysis tools, which include a large selection of algorithms developed to increase transparency in the decision-making process of Machine Learning (ML) models. These tools aim to provide users with interpretations of the data and the model. However, despite the abundance of options and their potential in identifying and decomposing model behavior, XAI's inability to quantitatively assess trustworthiness, due to the lack of quantifiable metrics, has resulted in low adoption in real-world applications. In contrast, traditional methods to evaluate trust such as uncertainty quantification, robust testing, and user studies scale well with large models and datasets, thanks to their reliance on quantifiable metrics. However, they do not offer the same level of transparency and qualitative assessments as XAI to make the models more interpretable, which are a key component of the multi-faceted trustworthiness assessment.</p><p dir="ltr">To bridge this gap, I propose a framework in which explanations produced by XAI are ensembled across a portfolio of models. These ensembled explanations are then used for both quantitative and qualitative comparison to evaluate trust in the models. The goal is to leverage these explanations to assess trustworthiness driven by transparency. The framework also identifies areas of consensus or disagreement among the ensembled explanations. Further leverage the presence or absence of consensus to bin model reasoning to indicate weaknesses, misalignment to user expectations, and/or distribution shifts.</p><p dir="ltr">A preliminary investigation of the proposed framework is carried out on multivariate time-series data from NASA's Commercial Modular Aero-Propulsion System Simulation (CMAPSS) to model and predict turbojet engine degradation. This approach uses three distinct ML models to forecast the remaining useful life (RUL) of the engine. Using the proposed framework, influential system parameters contributing to engine degradation in each model are identified via XAI. These explanations are ensembled and compared to assess consensus. Ultimately, the models disagree on the extent of certain features contributing to the failure. However, experimental literature supports this finding as modeling engine degradation can be sensitive to the type of failure mode. Additionally, certain model architectures work better for certain types of data patterns, leading to recommendations on expert models. With these results and understanding of the intricacies of the framework, it is revised and implemented on a more complex application with a different data type and task: defect detection in robotic manipulation. The ARMBench (Amazon Robotic Manipulation Benchmark) dataset is used to train computer vision models for an image-based multi-classification problem and explained using activation maps. In this use case, both upstream and downstream influences and benefits of the framework are assessed while assessing the trustworthiness of the model and its predictions. The framework throws light on the strengths and weaknesses of the models, dataset, and deployment. Aiding in identifying strategies to mitigate weak and untrustworthy models. </p>