Dissertations / Theses on the topic 'Réseaux neuronaux graphiques'
Create a spot-on reference in APA, MLA, Chicago, Harvard, and other styles
Consult the top 28 dissertations / theses for your research on the topic 'Réseaux neuronaux graphiques.'
Next to every source in the list of references, there is an 'Add to bibliography' button. Press on it, and we will generate automatically the bibliographic reference to the chosen work in the citation style you need: APA, MLA, Harvard, Chicago, Vancouver, etc.
You can also download the full text of the academic publication as pdf and read online its abstract whenever available in the metadata.
Browse dissertations / theses on a wide variety of disciplines and organise your bibliography correctly.
Lachaud, Guillaume. "Extensions and Applications of Graph Neural Networks." Electronic Thesis or Diss., Sorbonne université, 2023. http://www.theses.fr/2023SORUS434.
Full textGraphs are used everywhere to represent interactions between entities, whether physical such as atoms, molecules or people, or more abstract such as cities, friendships, ideas, etc. Amongst all the methods of machine learning that can be used, the recent advances in deep learning have made graph neural networks the de facto standard for graph representation learning. This thesis can be divided in two parts. First, we review the theoretical underpinnings of the most powerful graph neural networks. Second, we explore the challenges faced by the existing models when training on real world graph data. The powerfulness of a graph neural network is defined in terms of its expressiveness, i.e., its ability to distinguish non isomorphic graphs; or, in an equivalent manner, its ability to approximate permutation invariant and equivariant functions. We distinguish two broad families of the most powerful models. We summarise the mathematical properties as well as the advantages and disadvantages of these models in practical situations. Apart from the choice of the architecture, the quality of the graph data plays a crucial role in the ability to learn useful representations. Several challenges are faced by graph neural networks given the intrinsic nature of graph data. In contrast to typical machine learning methods that deal with tabular data, graph neural networks need to consider not only the features of the nodes but also the interconnectedness between them. Due to the connections between nodes, training neural networks on graphs can be done in two settings: in transductive learning, the model can have access to the test features in the training phase; in the inductive setting, the test data remains unseen. We study the differences in terms of performance between inductive and transductive learning for the node classification task. Additionally, the features that are fed to a model can be noisy or even missing. In this thesis we evaluate these challenges on real world datasets, and we propose a novel architecture to perform missing data imputation on graphs. Finally, while graphs can be the natural way to describe interactions, other types of data can benefit from being converted into graphs. In this thesis, we perform preliminary work on how to extract the most important parts of skin lesion images that could be used to create graphs and learn hidden relations in the data
Papanastasiou, Effrosyni. "Feasibility of Interactions and Network Inference of Online Social Networks." Electronic Thesis or Diss., Sorbonne université, 2024. http://www.theses.fr/2024SORUS173.
Full textThis thesis deals with the problem of network inference in the domain of Online So-cial Networks. The main premise of network inference problems is that the networkwe are observing is not the network that we really need. This is especially prevalentin today's digital space, where the abundance of information usually comes withcrucial unreliability, in the form of noise and missing points in the data. However, existing approaches either ignore or do not guarantee to infer networks in a waythat can explain the data we have at hand. As a result, there is an ambiguity around the meaning of the network that we are inferring, while also having little intuition or control over the inference itself. The goal of this thesis is to further explore this problem. To quantify how well an inferred network can explain a dataset, we introduce a novel quality criterion called feasibility. Our intuition is that if a dataset is feasible given an inferred network, we might also be closer to the ground truth. To verify this,we propose a novel network inference method in the form of a constrained, Maximum Likelihood-based optimization problem that guarantees 100% feasibility. It is tailored to inputs from Online Social Networks, which are well-known sources of un-reliable and restricted data. We provide extensive experiments on one synthetic andone real-world dataset coming from Twitter/X. We show that our proposed method generates a posterior distribution of graphs that guarantees to explain the dataset while also being closer to the true underlying structure when compared to other methods. As a final exploration, we look into the field of deep learning for more scalable and flexible alternatives, providing a preliminary framework based on Graph Neural Networks and contrastive learning that gives promising results
Badr, Bellaj. "Securing P2P resource sharing via blockchain and GNN-based trust." Electronic Thesis or Diss., Institut polytechnique de Paris, 2024. http://www.theses.fr/2024IPPAS005.
Full textThe emergence of blockchain technology and cryptocurrencies has enabled the development of innovative peer-to-peer (P2P) models for resource allocation, sharing, and monetization. As these P2P models operate without inherent trust, the need for reliable trust and reputation mechanisms becomes crucial to minimize potential risks associated with engaging with malicious peers. Several trust management systems (TMS) have been proposed to establish trust in traditional P2P networks, aiming to facilitate the selection of dependable resources and deter peer misbehavior, with a significant focus on utilizing reputation as a guiding factor.Reputation-based trust systems (RTMS) play a fundamental role by leveraging community-based reputations to establish trust. They enable peers to assess the trustworthiness of others and evaluate the Quality of Service (QoS) based on shared reputations and past interactions. While these systems establish a peer-to-peer overlay trust network, the majority of these protocols are not tailored to suit Blockchain-based networks, resulting in various shortcomings due to their outdated design.This thesis presents our protocol BTrust, a novel decentralized and modular trust management system for large-scale P2P networks, leveraging blockchain technology and (Graph Neural Network) GNN for trust evaluation. BTrust introduces a multi-dimensional trust and reputation model to assess peer trustworthiness, dynamically deriving a single value from multiple parameters. The blockchain ensures reliable trust computation, dissemination, and storage without a central trust manager.An important breakthrough in our protocol is the resolution of the "cold start" or "initial trust score problem". To achieve this, the bootstrapping peer adopts random walks to select trustworthy peers among its neighbors, ensuring a decentralized approach without relying on any centralized entity or predefined peers. Unlike existing solutions, this method prevents overwhelming the most trusted peers in the network.Another challenge addressed in reputation systems is the reluctance of peers to provide negative feedback, often due to fear of retaliation or simply not providing feedback at all. To tackle these issues, we introduce an incentive mechanism that encourages truthful feedback and implement specialized mechanisms to penalize bad or lazy behavior. These innovations promote a more reliable and balanced trust evaluation process within the system.Furthermore, we propose a variant of BTrust called GBTrust, which improves upon the original protocol by incorporating Graph Neural Networks (GNNs) and a novel attention-based mechanism specifically designed for trust management. This variant enhances the detection of dynamic malicious peers and strengthens the overall robustness and accuracy of trust evaluation. By leveraging GNNs, GBTrust effectively captures the complex relationships and dynamic behavior of peers in the network, enabling more accurate identification of malicious activities and better adaptability to changing trust dynamics. The attention-based mechanism further enhances the model's ability to prioritize and weigh different trust factors, leading to more reliable and precise trust assessments.We demonstrate the efficiency of the proposed protocol in large-scale P2P networks using simulations of a P2P network and show that BTrust and its variant (GBTrust) are highly resilient to failures and robust against malicious nodes
Hafidi, Hakim. "Robust machine learning for Graphs/Networks." Electronic Thesis or Diss., Institut polytechnique de Paris, 2023. http://www.theses.fr/2023IPPAT004.
Full textThis thesis addresses advancements in graph representation learning, focusing on the challengesand opportunities presented by Graph Neural Networks (GNNs). It highlights the significanceof graphs in representing complex systems and the necessity of learning node embeddings that capture both node features and graph structure. The study identifies key issues in GNNs, such as their dependence on high-quality labeled data, inconsistent performanceacross various datasets, and susceptibility to adversarial attacks.To tackle these challenges, the thesis introduces several innovative approaches. Firstly, it employs contrastive learning for node representation, enabling self-supervised learning that reduces reliance on labeled data. Secondly, a Bayesian-based classifier isproposed for node classification, which considers the graph’s structure to enhance accuracy. Lastly, the thesis addresses the vulnerability of GNNs to adversarialattacks by assessing the robustness of the proposed classifier and introducing effective defense mechanisms.These contributions aim to improve both the performance and resilience of GNNs in graph representation learning
Akakzia, Ahmed. "Teaching Predicate-based Autotelic Agents." Electronic Thesis or Diss., Sorbonne université, 2022. https://accesdistant.sorbonne-universite.fr/login?url=https://theses-intra.sorbonne-universite.fr/2022SORUS415.pdf.
Full textAs part of the quest for designing embodied machines that autonomously explore their environments, discover new behaviors and acquire open-ended repertoire of skills, artificial intelligence has been taking long looks at the inspiring fields of developmental psychology and cognitive sciences which investigate the remarkable continuous and unbounded learning of humans. This gave birth to the field of developmental robotics which aims at designing autonomous artificial agents capable of self-organizing their own learning trajectories based on their intrinsic motivations. It bakes the developmental framework of intrinsically motivated goal exploration processes (IMGEPs) into reinforcement learning (RL). This combination has been recently introduced as autotelic reinforcement learning, where autotelic agents are intrinsically motivated to self-represent, self-organize and autonomously learn about their own goals. Naturally, such agents need to be endowed with good exploration capabilities as they need to first physically encounter a certain goal in order to take ownership of and learn about it. Unfortunately, discovering interesting behavior is usually tricky, especially in hard exploration setups where the rewarding signals are parsimonious, deceptive or adversarial. In such scenarios, the agents’ physical situatedness-in the Piagetian sense of the term-seems insufficient. Luckily, research in developmental psychology and education sciences have been praising the remarkable role of socio-cultural signals in the development of human children. This social situatedness-in the Vygotskyan sense of the term-enhances the toddlers’ exploration capabilities, creativity and development. However, deep \rl considers social interactions as dictating instructions to the agents, depriving them from their autonomy. This research introduces \textit{teachable autotelic agents}, a novel family of autonomous machines that can learn both alone and from external social signals. We formalize such a family as a hybrid goal exploration process (HGEPs), where autotelic agents are endowed with an internalization mechanism to rehearse social signals and with a goal source selector to actively query for social guidance. The present manuscript is organized in two parts. In the first part, we focus on the design of teachable autotelic agents and attempt to leverage the most important properties that would later serve the social interaction. Namely, we introduce predicate-based autotelic agents, a novel family of autotelic agents that represent their goals using spatial binary predicates. These insights were based on the Mandlerian view on the prelinguistic concept acquisition suggesting that toddlers are endowed with some innate mechanisms enabling them to translate spatio-temporal information into an iconic static form. We show that the underlying semantic representation plays a pivotal role between raw sensory inputs and language inputs, enabling the decoupling of sensorimotor learning and language grounding. We also investigate the design of such agents' policies and state-action value functions, and argue that combining Graph Neural Networks (GNNs) with relational predicates provides a light computational scheme to transfer efficiently between skills. In the second part, we formalize social interactions as a goal exploration process. We introduce Help Me Explore (HME), a novel social interaction protocol where an expert social partner progressively guides the learning agent beyond its zone of proximal development (ZPD). The agent actively selects to query its social partner whenever it estimates that it is not progressing enough alone. It eventually internalizes the social signals, becomes less dependent on its social partner and maximizes its control over its goal space
Wang, Lianfa. "Improving the confidence of CFD results by deep learning." Electronic Thesis or Diss., Université Paris sciences et lettres, 2024. http://www.theses.fr/2024UPSLM008.
Full textComputational Fluid Dynamics (CFD) has become an indispensable tool for studying complex flow phenomena in both research and industry over the years. The accuracy of CFD simulations depends on various parameters – geometry, mesh, schemes, solvers, etc. – as well as phenomenological knowledge that only an expert CFD engineer can configure and optimize. The objective of this thesis is to propose an AI assistant to help users, whether they are experts or not, to better choose simulation options and ensure the reliability of results for a target flow phenomenon. In this context, deep learning algorithms are explored to identify the characteristics of flows computed on structured and unstructured meshes of complex geometries. Initially, convolutional neural networks (CNNs), known for their ability to extract patterns from im-ages, are used to identify flow phenomena such as vortices and thermal stratification on structured 2D meshes. Although the results obtained on structured meshes are satisfactory, CNNs can only be applied to structured meshes. To overcome this limitation, a graph-based neural network (GNN) framework is proposed. This framework uses the U-Net architecture and a hierarchy of successively refined graphs through the implementation of a multigrid method (AMG) inspired by the one used in the Code_Saturne CFD code. Subsequently, an in-depth study of kernel functions was conducted according to identification accuracy and training efficiency criteria to better filter the different phenomena on unstructured meshes. After comparing available kernel functions in the literature, a new kernel function based on the Gaussian mixture model was proposed. This function is better suited to identifying flow phenomena on unstructured meshes. The superiority of the proposed architecture and kernel function is demonstrated by several numerical experiments identifying 2D vortices and its adaptability to identifying the characteristics of a 3D flow
Khessiba, Souhir. "Stratégies d’optimisation des hyper-paramètres de réseaux de neurones appliqués aux signaux temporels biomédicaux." Electronic Thesis or Diss., Institut polytechnique de Paris, 2024. http://www.theses.fr/2024IPPAE003.
Full textThis thesis focuses on optimizing the hyperparameters of convolutional neural networks (CNNs) in the medical domain, proposing an innovative approach to improve the performance of decision-making models in the biomedical field. Through the use of a hybrid approach, GS-TPE, to effectively adjust the hyperparameters of complex neural network models, this research has demonstrated significant improvements in the classification of temporal biomedical signals, such as vigilance states, from physiological signals such as electroencephalogram (EEG). Furthermore, by introducing a new DNN architecture, STGCN, for the classification of gestures associated with pathologies such as knee osteoarthritis and Parkinson's disease from video gait analysis, these works offer new perspectives for enhancing medical diagnosis and management through advancements in artificial intelligence
Prouteau, Thibault. "Graphs,Words, and Communities : converging paths to interpretability with a frugal embedding framework." Electronic Thesis or Diss., Le Mans, 2024. http://www.theses.fr/2024LEMA1006.
Full textRepresentation learning with word and graph embedding models allows distributed representations of information that can in turn be used in input of machine learning algorithms. Through the last two decades, the tasks of embedding graphs’ nodes and words have shifted from matrix factorization approaches that could be trained in a matter of minutes to large models requiring ever larger quantities of training data and sometimes weeks on large hardware architectures. However, in a context of global warming where sustainability is a critical concern, we ought to look back to previous approaches and consider their performances with regard to resources consumption. Furthermore, with the growing involvement of embeddings in sensitive machine learning applications (judiciary system, health), the need for more interpretable and explainable representations has manifested. To foster efficient representation learning and interpretability, this thesis introduces Lower Dimension Bipartite Graph Framework (LDBGF), a node embedding framework able to embed with the same pipeline graph data and text from large corpora represented as co-occurrence networks. Within this framework, we introduce two implementations (SINr-NR, SINr-MF) that leverage community detection in networks to uncover a latent embedding space where items (nodes/words) are represented according to their links to communities. We show that SINr-NR and SINr-MF can compete with similar embedding approaches on tasks such as predicting missing links in networks (link prediction) or node features (degree centrality, PageRank score). Regarding word embeddings, we show that SINr-NR is a good contender to represent words via word co-occurrence networks. Finally, we demonstrate the interpretability of SINr-NR on multiple aspects. First with a human evaluation that shows that SINr-NR’s dimensions are to some extent interpretable. Secondly, by investigating sparsity of vectors, and how having fewer dimensions may allow interpreting how the dimensions combine and allow sense to emerge
Liu, Wenzhuo. "Deep Graph Neural Networks for Numerical Simulation of PDEs." Electronic Thesis or Diss., université Paris-Saclay, 2023. http://www.theses.fr/2023UPASG032.
Full textPartial differential equations (PDEs) are an essential modeling tool for the numerical simulation of complex systems. However, their accurate numerical resolution usually requires a high computational cost. In recent years, deep Learning algorithms have demonstrated impressive successes in learning from examples, and their direct application to databases of existing solutions of a PDE could be a way to tackle the excessive computational cost of classical numerical approaches: Once a neural model has been learned, the computational cost of inference of the solution on new example is very low. However, many issues remain that this Ph.D. thesis investigates, focusing on three major hurdles: handling unstructured meshes, which can hardly be done accurately by simply porting the neural successes on image processing tasks; generalization issues, in particular for Out-of-Distribution examples; and the too high computational costs for generating the training data. We propose three contributions, based on Graph Neural Networks, to tackle these problems: A hierarchical model inspired by the multi-grid techniques of Numerical Analysis; The use of Meta-Learning to improve the performance of Out-of-Distribution data; and Transfer Learning between multi-fidelity datasets to reduce the computational cost of data generation. The proposed approaches are experimentally validated on different physical systems
Halal, Taha. "Graph-based learning and optimization." Electronic Thesis or Diss., université Paris-Saclay, 2024. http://www.theses.fr/2024UPASG043.
Full textGraphs are a fundamental data structure used to represent complex patterns in various domains. Graph Neural Networks (GNNs), a deep learning paradigm specifically designed for graph-structured data, offer a powerful deep learning solution for extracting insights from these intricate relationships. This thesis explores the application of GNNs to address two key challenges: maximizing influence in social networks and predicting missing links in knowledge graphs with limited data. With applications ranging from optimizing public health campaigns and combating misinformation to knowledge base completion, this research addresses the need for computationally efficient and robust methods in these domains. Influence maximization (IM) focuses on identifying the most influential nodes within a social network to maximize the spread of information or ideas. This thesis explores methods for tackling the IM problem, particularly in real-world scenarios with massive networks and diverse information themes. We build our models upon the S2V-DQN framework, a powerful approach that combines Deep Q-Networks (DQNs) for reinforcement learning with Structure2Vec (S2V) for graph embedding. We first develop our IM-GNN model that incorporates advanced GNN features such as graph attention mechanisms and positional encoding, demonstrating competitive performance against existing learning-based and non-learning based methods for influence maximization. We further extend our research to tackle Topic-aware Influence Maximization (TIM) where the spread of information is influenced by its thematic content, requiring models to consider not only network structure but also the topics of the messages being shared. This is where the limitations of traditional IM methods become apparent. Our TIM-GNN model effectively handles this complexity by incorporating topic-aware training and probabilistic methods for constructing topic-aware diffusion graphs. To address query latency concerns, we introduce TIM-GNNx, which integrates cross-attention mechanisms and a pre-computed Q-matrix. Our experiments on real-world datasets demonstrate that our proposed model achieves competitive performance in terms of influence spread compared to state-of-the-art methods while also offering significant improvements in query time latency and robustness to changes in the diffusion graph. Notably, our TIM-GNNx model strikes a balance between query efficiency and maximizing influence, making it particularly well-suited for real-time applications. In the realm of knowledge graphs, we explore Few-Shot Link Prediction (FSLP), where the goal is to predict missing relationships with limited training examples, which is crucial for addressing the long-tail phenomenon. In knowledge graphs, the long-tail phenomenon refers to the fact that a large number of entities (nodes) and relations (edges) have very few connections or occurrences. This results in a distribution where a small number of popular entities or relations have many connections, while the vast majority have very few. Our investigation focuses on the feasibility of integrating a path-based knowledge graph completion method PathCon with a meta-learning framework MetaR to address the limitations of the latter. While our initial investigations did not yield significant improvements or notable scientific contributions, they provided valuable insights into the challenges of this task and informed the development of a prototype, deployed as an API, for the AIDA project. This prototype demonstrates the practical value of our research and paves the way for future explorations in this area. Overall, this thesis contributes novel and efficient GNN-based solutions for influence maximization and explores promising directions for few-shot link prediction in knowledge graphs, pushing the boundaries of these research areas
Sourty, Raphael. "Apprentissage de représentation de graphes de connaissances et enrichissement de modèles de langue pré-entraînés par les graphes de connaissances : approches basées sur les modèles de distillation." Electronic Thesis or Diss., Toulouse 3, 2023. http://www.theses.fr/2023TOU30337.
Full textNatural language processing (NLP) is a rapidly growing field focusing on developing algorithms and systems to understand and manipulate natural language data. The ability to effectively process and analyze natural language data has become increasingly important in recent years as the volume of textual data generated by individuals, organizations, and society as a whole continues to grow significantly. One of the main challenges in NLP is the ability to represent and process knowledge about the world. Knowledge graphs are structures that encode information about entities and the relationships between them, they are a powerful tool that allows to represent knowledge in a structured and formalized way, and provide a holistic understanding of the underlying concepts and their relationships. The ability to learn knowledge graph representations has the potential to transform NLP and other domains that rely on large amounts of structured data. The work conducted in this thesis aims to explore the concept of knowledge distillation and, more specifically, mutual learning for learning distinct and complementary space representations. Our first contribution is proposing a new framework for learning entities and relations on multiple knowledge bases called KD-MKB. The key objective of multi-graph representation learning is to empower the entity and relation models with different graph contexts that potentially bridge distinct semantic contexts. Our approach is based on the theoretical framework of knowledge distillation and mutual learning. It allows for efficient knowledge transfer between KBs while preserving the relational structure of each knowledge graph. We formalize entity and relation inference between KBs as a distillation loss over posterior probability distributions on aligned knowledge. Grounded on this finding, we propose and formalize a cooperative distillation framework where a set of KB models are jointly learned by using hard labels from their own context and soft labels provided by peers. Our second contribution is a method for incorporating rich entity information from knowledge bases into pre-trained language models (PLM). We propose an original cooperative knowledge distillation framework to align the masked language modeling pre-training task of language models and the link prediction objective of KB embedding models. By leveraging the information encoded in knowledge bases, our proposed approach provides a new direction to improve the ability of PLM-based slot-filling systems to handle entities
Ferré, Paul. "Adéquation algorithme-architecture de réseaux de neurones à spikes pour les architectures matérielles massivement parallèles." Thesis, Toulouse 3, 2018. http://www.theses.fr/2018TOU30318/document.
Full textThe last decade has seen the re-emergence of machine learning methods based on formal neural networks under the name of deep learning. Although these methods have enabled a major breakthrough in machine learning, several obstacles to the possibility of industrializing these methods persist, notably the need to collect and label a very large amount of data as well as the computing power necessary to perform learning and inference with this type of neural network. In this thesis, we propose to study the adequacy between inference and learning algorithms derived from biological neural networks and massively parallel hardware architectures. We show with three contribution that such adequacy drastically accelerates computation times inherent to neural networks. In our first axis, we study the adequacy of the BCVision software engine developed by Brainchip SAS for GPU platforms. We also propose the introduction of a coarse-to-fine architecture based on complex cells. We show that GPU portage accelerates processing by a factor of seven, while the coarse-to-fine architecture reaches a factor of one thousand. The second contribution presents three algorithms for spike propagation adapted to parallel architectures. We study exhaustively the computational models of these algorithms, allowing the selection or design of the hardware system adapted to the parameters of the desired network. In our third axis we present a method to apply the Spike-Timing-Dependent-Plasticity rule to image data in order to learn visual representations in an unsupervised manner. We show that our approach allows the effective learning a hierarchy of representations relevant to image classification issues, while requiring ten times less data than other approaches in the literature
Giraldo, Zuluaga Jhony Heriberto. "Graph-based Algorithms in Computer Vision, Machine Learning, and Signal Processing." Electronic Thesis or Diss., La Rochelle, 2022. http://www.theses.fr/2022LAROS037.
Full textGraph representation learning and its applications have gained significant attention in recent years. Notably, Graph Neural Networks (GNNs) and Graph Signal Processing (GSP) have been extensively studied. GNNs extend the concepts of convolutional neural networks to non-Euclidean data modeled as graphs. Similarly, GSP extends the concepts of classical digital signal processing to signals supported on graphs. GNNs and GSP have numerous applications such as semi-supervised learning, point cloud semantic segmentation, prediction of individual relations in social networks, modeling proteins for drug discovery, image, and video processing. In this thesis, we propose novel approaches in video and image processing, GNNs, and recovery of time-varying graph signals. Our main motivation is to use the geometrical information that we can capture from the data to avoid data hungry methods, i.e., learning with minimal supervision. All our contributions rely heavily on the developments of GSP and spectral graph theory. In particular, the sampling and reconstruction theory of graph signals play a central role in this thesis. The main contributions of this thesis are summarized as follows: 1) we propose new algorithms for moving object segmentation using concepts of GSP and GNNs, 2) we propose a new algorithm for weakly-supervised semantic segmentation using hypergraph neural networks, 3) we propose and analyze GNNs using concepts from GSP and spectral graph theory, and 4) we introduce a novel algorithm based on the extension of a Sobolev smoothness function for the reconstruction of time-varying graph signals from discrete samples
Tiano, Donato. "Learning models on healthcare data with quality indicators." Electronic Thesis or Diss., Lyon 1, 2022. http://www.theses.fr/2022LYO10182.
Full textTime series are collections of data obtained through measurements over time. The purpose of this data is to provide food for thought for event extraction and to represent them in an understandable pattern for later use. The whole process of discovering and extracting patterns from the dataset is carried out with several extraction techniques, including machine learning, statistics, and clustering. This domain is then divided by the number of sources adopted to monitor a phenomenon. Univariate time series when the data source is single and multivariate time series when the data source is multiple. The time series is not a simple structure. Each observation in the series has a strong relationship with the other observations. This interrelationship is the main characteristic of time series, and any time series extraction operation has to deal with it. The solution adopted to manage the interrelationship is related to the extraction operations. The main problem with these techniques is that they do not adopt any pre-processing operation on the time series. Raw time series have many undesirable effects, such as noisy points or the huge memory space required for long series. We propose new data mining techniques based on the adoption of the most representative features of time series to obtain new models from the data. The adoption of features has a profound impact on the scalability of systems. Indeed, the extraction of a feature from the time series allows for the reduction of an entire series to a single value. Therefore, it allows for improving the management of time series, reducing the complexity of solutions in terms of time and space. FeatTS proposes a clustering method for univariate time series that extracts the most representative features of the series. FeatTS aims to adopt the features by converting them into graph networks to extract interrelationships between signals. A co-occurrence matrix merges all detected communities. The intuition is that if two time series are similar, they often belong to the same community, and the co-occurrence matrix reveals this. In Time2Feat, we create a new multivariate time series clustering. Time2Feat offers two different extractions to improve the quality of the features. The first type of extraction is called Intra-Signal Features Extraction and allows to obtain of features from each signal of the multivariate time series. Inter-Signal Features Extraction is used to obtain features by considering pairs of signals belonging to the same multivariate time series. Both methods provide interpretable features, which makes further analysis possible. The whole time series clustering process is lighter, which reduces the time needed to obtain the final cluster. Both solutions represent the state of the art in their field. In AnomalyFeat, we propose an algorithm to reveal anomalies from univariate time series. The characteristic of this algorithm is the ability to work among online time series, i.e. each value of the series is obtained in streaming. In the continuity of previous solutions, we adopt the functionality of revealing anomalies in the series. With AnomalyFeat, we unify the two most popular algorithms for anomaly detection: clustering and recurrent neural network. We seek to discover the density area of the new point obtained with clustering
Mekemeza, Ona Keshia. "Photonic spiking neuron network." Electronic Thesis or Diss., Bourgogne Franche-Comté, 2023. http://www.theses.fr/2023UBFCD052.
Full textToday, neuromorphic networks play a crucial role in information processing,particularly as tasks become increasingly complex: voice recognition, dynamic image correlation, rapid multidimensional decision- making, data merging, behavioral optimization, etc... Neuromorphic networks come in several types; spiking networks are one of them. The latter's modus operandi is based on that of cortical neurons. As spiking networks are the most energy-efficient neuromorphic networks, they offer the greatest potential for scaling. Several demonstrations of artificial neurons have been conducted with electronic and more recently photonic circuits. The integration density of silicon photonics is an asset to create circuits that are complex enough to hopefully carry out a complete demonstration. Therefore, this thesis aims to exploit an architecture of a photonic spiking neural network based on Q-switched lasers integrated into silicon and an ultra-dense and reconfigurable interconnection circuit that can simulate synaptic weights. A complete modeling of the circuit is expected with a practical demonstration of an application in solving a mathematical problem to be defined
Boschin, Armand. "Machine learning techniques for automatic knowledge graph completion." Electronic Thesis or Diss., Institut polytechnique de Paris, 2023. http://www.theses.fr/2023IPPAT016.
Full textA knowledge graph is a directed graph in which nodes are entities and edges, typed by a relation, represent known facts linking two entities. These graphs can encode a wide variety of information, but their construction and exploitation can be complex. Historically, symbolic methods have been used to extract rules about entities and relations, to correct anomalies or to predict missing facts. More recently, techniques of representation learning, or embeddings, have attempted to solve these same tasks. Initially purely algebraic or geometric, these methods have become more complex with deep neural networks and have sometimes been combined with pre-existing symbolic techniques.In this thesis, we first focus on the problem of implementation. Indeed, the diversity of libraries used makes the comparison of results obtained by different models a complex task. In this context, the Python library TorchKGE was developed to provide a unique setup for the implementation of embedding models and a highly efficient inference evaluation module. This library relies on graphic acceleration of tensor computation provided by PyTorch, is compatible with widespread optimization libraries and is available as open source.We then consider the automatic enrichment of Wikidata by typing the hyperlinks linking Wikipedia pages. A preliminary study showed that the graph of Wikipedia articles is much denser than the corresponding knowledge graph in Wikidata. A new training method involving relations and an inference method using entity types were proposed and experiments showed the relevance of the combined approach, including on a new dataset.Finally, we explore automatic entity typing as a hierarchical classification task. That led to the design of a new hierarchical loss used to train tensor-based models along with a new type of encoder. Experiments on two datasets have allowed a good understanding of the impact a prior knowledge of class taxonomy can have on a classifier but also reinforced the intuition that the hierarchy can be learned from the features if the dataset is large enough
Anakok, Emre. "Prise en compte des effets d'échantillonnage pour la détection de structure des réseaux écologiques." Electronic Thesis or Diss., université Paris-Saclay, 2024. http://www.theses.fr/2024UPASM049.
Full textIn this thesis, we focus on the biases that sampling can cause on the estimation of statistical models and metrics describing ecological interaction networks. First, we propose to combine an observation model that accounts for sampling with a stochastic block model representing the structure of possible interactions. The identifiability of the model is demonstrated and an algorithm is proposed to estimate its parameters. Its relevance and its practical interest are attested on a large dataset of plant-pollinator networks, as we observe structural change on most of the networks. We then examine a large dataset sampled by a citizen science program. Using recent advances in artificial intelligence, we propose a method to reconstruct the ecological network free from sampling effects caused by the varying levels of experience among observers. Finally, we present methods to highlight variables of ecological interest that influence the network's connectivity and show that accounting for sampling effects partially alters the estimation of these effects. Our methods, implemented in either R or Python, are freely accessible
Hladiš, Matej. "Réseaux de neurones en graphes et modèle de langage des protéines pour révéler le code combinatoire de l'olfaction." Electronic Thesis or Diss., Université Côte d'Azur, 2024. http://www.theses.fr/2024COAZ5024.
Full textMammals identify and interpret a myriad of olfactory stimuli using a complex coding mechanism involving interactions between odorant molecules and hundreds of olfactory receptors (ORs). These interactions generate unique combinations of activated receptors, called the combinatorial code, which the human brain interprets as the sensation we call smell. Until now, the vast number of possible receptor-molecule combinations have prevented a large-scale experimental study of this code and its link to odor perception. Therefore, revealing this code is crucial to answering the long-term question of how we perceive our intricate chemical environment. ORs belong to the class A of G protein-coupled receptors (GPCRs) and constitute the largest known multigene family. To systematically study olfactory coding, we develop M2OR, a comprehensive database compiling the last 25 years of OR bioassays. Using this dataset, a tailored deep learning model is designed and trained. It combines the [CLS] token embedding from a protein language model with graph neural networks and multi-head attention. This model predicts the activation of ORs by odorants and reveals the resulting combinatorial code for any odorous molecule. This approach is refined by developing a novel model capable of predicting the activity of an odorant at a specific concentration, subsequently allowing the estimation of the EC50 value for any OR-odorant pair. Finally, the combinatorial codes derived from both models are used to predict the odor perception of molecules. By incorporating inductive biases inspired by olfactory coding theory, a machine learning model based on these codes outperforms the current state-of-the-art in smell prediction. To the best of our knowledge, this is the most comprehensive and successful application of combinatorial coding to odor quality prediction. Overall, this work provides a link between the complex molecule-receptor interactions and human perception
Hu, Xu. "Towards efficient learning of graphical models and neural networks with variational techniques." Thesis, Paris Est, 2019. http://www.theses.fr/2019PESC1037.
Full textIn this thesis, I will mainly focus on variational inference and probabilistic models. In particular, I will cover several projects I have been working on during my PhD about improving the efficiency of AI/ML systems with variational techniques. The thesis consists of two parts. In the first part, the computational efficiency of probabilistic graphical models is studied. In the second part, several problems of learning deep neural networks are investigated, which are related to either energy efficiency or sample efficiency
Nastorg, Matthieu. "Scalable GNN Solutions for CFD Simulations." Electronic Thesis or Diss., université Paris-Saclay, 2024. http://www.theses.fr/2024UPASG020.
Full textComputational Fluid Dynamics (CFD) plays an essential role in predicting various physical phenomena, such as climate, aerodynamics, or blood flow. At the core of CFD lie the Navier-Stokes equations governing the motion of fluids. However, solving these equations at scale remains daunting, especially when dealing with Incompressible Navier-Stokes equations. Indeed, the well-known splitting schemes require the costly resolution of a Pressure Poisson problem that guarantees the incompressibility constraint. Nowadays, Deep Learning methods have opened new perspectives for enhancing numerical simulations. Among existing approaches, Graph Neural Networks (GNNs), designed to handle graph data like meshes, have proven to be promising. This thesis is dedicated to exploring the use of GNNs to enhance the resolution of the Pressure Poisson problem. One significant contribution involves introducing a novel physics-informed GNN-based model that inherently respects boundary conditions while leveraging the Implicit Layer theory to automatically adjust the number of GNN layers required for convergence. This results in a model with enhanced generalization capabilities, effectively handling Poisson problems of various sizes and shapes. Nevertheless, its current limitations restrict it to small-scale problems, insufficient for industrial applications that often require thousands of nodes. To scale up these models, this thesis further explores combining GNNs with Domain Decomposition methods, taking advantage of batch parallel computing on GPU to produce more efficient engineering solutions
Tardif, Malo. "Proximal sensing and neural network processes to assist in diagnosis of multi-symptom grapevine diseases." Electronic Thesis or Diss., Bordeaux, 2023. http://www.theses.fr/2023BORD0369.
Full textGrapevine is a plant susceptible to numerous diseases. Some of these diseases can lead to significant yield losses and the death of the infected grapevine. Among these diseases, some present symptoms of different nature on various organs of the same vine. Their diagnosis, typically performed by experts, is even more complex as many confounding factors are present. This research focuses on the development of methodologies for acquiring, annotating, and processing data related to multi-symptom grapevine diseases to study their automated diagnosis. Two groups of diseases are targeted: grapevine yellows such as Flavescence dorée (FD) and grapevine trunk diseases (GTDs) with Eutypa and Botryosphaeria diebacks as specific diseases.RGB image acquisitions were conducted directly in grapevine rows to build datasets for each disease type. The dataset for FD covers five different grape varieties and takes into account many diseases that have symptoms similar to FD, referred to as confounding diseases. The GTDs dataset includes images of a single grape variety and no confounding disease. Three methods for the automatic diagnosis of these diseases are proposed, compared, and discussed. The first method, inspired by state-of-the-art techniques, uses a convolutional neural network-based classifier applied to raw images (method A). The results show that this methodology delivers good results on datasets containing very few confounding diseases. Precision (p) and recall (r) of (p=0.94, r=0.92) are achieved for classifying images of grapevines affected by GTDs, while they are (p=0.87, r=0.84) for classifying images of vines affected by FD in a dataset containing 16% of confounding disease images.To improve these results, two methods were developed, both consisting of two steps: (1) individual symptom detection using a detection algorithm composed of neural convolutional layers and a neural segmentation algorithm; (2) diagnosis based on the association of detected symptoms, either using a Random Forest classifier (method B) or a graph neural network (method C). The results of these two methodologies on the dataset containing 16% of confounding disease images for FD are (p=0.86, r=0.90) for method B and (p=0.90, r=0.96) for method C. These results demonstrate the better effectiveness of two-step methodologies in distinguishing confounding diseases from targeted diseases. They also highlight the relevance of embedded RGB imaging combined with artificial intelligence approaches for diagnosing these diseases.Finally, these three methods are tested on whole-block acquisitions to establish their validity in real-world use cases. The results highlight the advantages of the two-step methodology based on symptom association by graph, the significant contribution of considering the surrounding vines and both sides of the vines during their automated diagnosis, and emphasize the challenges of real-world application of these methodologies
Seznec, Mickaël. "From the algorithm to the targets, optimization flow for high performance computing on embedded GPUs." Electronic Thesis or Diss., université Paris-Saclay, 2021. http://www.theses.fr/2021UPASG074.
Full textCurrent digital processing algorithms require more computing power to achieve more accurate results and process larger data. In the meantime, hardware architectures are becoming more specialized, with highly efficient accelerators designed for specific tasks. In this context, the path of deployment from the algorithm to the implementation becomes increasingly complex. It is, therefore, crucial to determine how algorithms can be modified to take advantage of new hardware capabilities. Our study focused on graphics processing units (GPUs), a massively parallel processor. Our algorithmic work was done in the context of radio-astronomy or optical flow estimation and consisted of finding the best adaptation of the software to the hardware. At the level of a mathematical operator, we modified the traditional image convolution algorithm to use the matrix units and showed that its performance doubles for large convolution kernels. At a broader method level, we evaluated linear solvers for the combined local-global optical flow to find the most suitable one on GPU. With additional optimizations, such as iteration fusion or memory buffer re-utilization, the method is twice as fast as the initial implementation, running at 60 frames per second on an embedded platform (30 W). Finally, we also pointed out the interest of this hardware-aware algorithm design method in the context of deep neural networks. For that, we showed the hybridization of a convolutional neural network for optical flow estimation with a pre-trained image classification network, MobileNet, that was initially designed for efficient image classification on low-power platforms
Guesdon, Romain. "Estimation de poses humaines par apprentissage profond : application aux passagers des véhicules autonomes." Electronic Thesis or Diss., Lyon 2, 2024. http://www.theses.fr/2024LYO20002.
Full textResearch into autonomous cars has made great strides in recent decades, focusing particularly on analysis of the external environment and driving-related tasks. This has led to a significant increase in the autonomy of private vehicles. In this new context, it may be relevant to take an interest in the passengers of these autonomous vehicles, to study their behavior in the face of this revolution in the means of transport. The AURA AutoBehave project has been set up to explore these issues in greater depth. This project brings together several laboratories conducting research in different scientific disciplines linked to this theme, such as computer vision, biomechanics, emotions, and transport economics. This thesis carried out at the LIRIS laboratory is part of this project, in which we focus on methods for estimating the human poses of passengers using deep learning. We first looked at state-of-the-art solutions and developed both a dataset and a metric better suited to the constraints of our context. We also studied the visibility of the keypoints to help estimate the pose. We then tackled the problem of domain generalisation for pose estimation to propose an efficient solution under unknown conditions. Thus, we focused on the generation of synthetic passenger data for pose estimation. Among other things, we studied the application of generative networks and 3D modeling methods to our problem. We have used this data to propose different training strategies and two new network architectures. The proposed fusion approach associated with the training strategies makes it possible to take advantage of both generic and specific datasets, to improve the generalisation capabilities of pose estimation methods inside a car, particularly on the lower body
Hubert, Nicolas. "Mesure et enrichissement sémantiques des modèles à base d'embeddings pour la prédiction de liens dans les graphes de connaissances." Electronic Thesis or Diss., Université de Lorraine, 2024. http://www.theses.fr/2024LORR0059.
Full textKnowledge graph embedding models (KGEMs) have gained considerable traction in recent years. These models learn a vector representation of knowledge graph entities and relations, a.k.a. knowledge graph embeddings (KGEs). This thesis specifically explores the advancement of KGEMs for the link prediction (LP) task, which is of utmost importance as it underpins several downstream applications such as recommender systems. In this thesis, various challenges around the use of KGEMs for LP are identified: the scarcity of semantically rich resources, the unidimensional nature of evaluation frameworks, and the lack of semantic considerations in prevailing machine learning-based approaches. Central to this thesis is the proposition of novel solutions to these challenges. Firstly, the thesis contributes to the development of semantically rich resources: mainstream datasets for link prediction are enriched using schema-based information, EducOnto and EduKG are proposed to overcome the paucity of resources in the educational domain, and PyGraft is introduced as an innovative open-source tool for generating synthetic ontologies and knowledge graphs. Secondly, the thesis proposes a new semantic-oriented evaluation metric, Sem@K, offering a multi-dimensional perspective on model performance. Importantly, popular models are reassessed using Sem@K, which reveals essential insights into their respective capabilities and highlights the need for multi-faceted evaluation frameworks. Thirdly, the thesis delves into the development of neuro-symbolic approaches, transcending traditional machine learning paradigms. These approaches do not only demonstrate improved semantic awareness but also extend their utility to diverse applications such as recommender systems. In summary, the present work not only redefines the evaluation and functionality of knowledge graph embedding models but also sets the stage for more versatile, interpretable AI systems, underpinning future explorations at the intersection of machine learning and symbolic reasoning
Meyer, Lucas. "Deep Learning en Ligne pour la Simulation Numérique à Grande Échelle." Electronic Thesis or Diss., Université Grenoble Alpes, 2024. http://www.theses.fr/2024GRALM001.
Full textMany engineering applications and scientific discoveries rely on faithful numerical simulations of complex phenomena. These phenomena are transcribed mathematically into Partial Differential Equation (PDE), whose solution is generally approximated by solvers that perform intensive computation and generate tremendous amounts of data. The applications rarely require only one simulation but rather a large ensemble of runs for different parameters to analyze the sensitivity of the phenomenon or to find an optimal configuration. Those large ensemble runs are limited by computation time and finite memory capacity. The high computational cost has led to the development of high-performance computing (HPC) and surrogate models. Recently, pushed by the success of deep learning in computer vision and natural language processing, the scientific community has considered its use to accelerate numerical simulations. The present thesis follows this approach by first presenting two techniques using machine learning for surrogate models. First, we propose to use a series of convolutions on hierarchical graphs to reproduce the velocity of fluids as generated by solvers at any time of the simulation. Second, we hybridize regression algorithms with classical reduced-order modeling techniques to identify the coefficients of any new simulation in a reduced basis computed by proper orthogonal decomposition. These two approaches, as the majority found in the literature, are supervised. Their training needs to generate a large number of simulations. Thus, they suffer the same problem that motivated their development in the first instance: generating many faithful simulations at scale is laborious. We propose a generic training framework for artificial neural networks that generate data simulations on-the-fly by leveraging HPC resources. Data are produced by running simultaneously several instances of the solver for different parameters. The solver itself can be parallelized over several processing units. As soon as a time step is computed by any simulation, it is streamed for training. No data is ever written on disk, thus overcoming slow input-output operations and alleviating the memory footprint. Training is performed by several GPUs with distributed data-parallelism. Because the training is now online, it induces a bias in the data compared to classical training, for which they are sampled uniformly from an ensemble of simulations available a priori. To mitigate this bias, each GPU is associated with a memory buffer in charge of mixing the incoming simulation data. This framework has improved the generalization capabilities of state-of-the-art architectures by exposing them during training to a richer diversity of data than would have been feasible with classical training. Experiments show the importance of the memory buffer implementation in guaranteeing generalization capabilities and high throughput training. The framework has been used to train a deep surrogate for heat diffusion simulation in less than 2 hours on 8TB of data processed in situ, thus increasing the prediction accuracy by 47% compared to a classical setting
Maiboroda, Vera. "Measurement of the differential cross-section of the ttH process and its CP properties in the multilepton final state using deep learning with the ATLAS experiment." Electronic Thesis or Diss., université Paris-Saclay, 2024. http://www.theses.fr/2024UPASP062.
Full textThis thesis measures the ttH production cross-section and the CP properties of the top Yukawa coupling within the Simplified Template Cross-Sections (STXS) framework. This study is performed in the multilepton final state (using two leptons with same sign and three leptons) with the full 140fb⁻¹ dataset recorded by the ATLAS detector at the LHC. Implementing the STXS framework requires reconstructing the Higgs transverse momentum, which is challenging due to the presence of several neutrinos in the final states. A Graph Neural Network approach was explored for this reconstruction. As the Higgs transverse momentum is also sensitive to the CP properties of the top Yukawa coupling, such reconstruction helps to study CP violation in the ttH process.The ttH STXS cross-section is measured with an expected uncertainty ranging from: -0.85 to 1.15. The expected limit on the CP-mixing angle that drives the amount CP violation in the top Yukawa coupling is: |α/ π| < 0.31 at 95% CL. For the high-luminosity phase of the LHC, the ATLAS tracker will be replaced with an upgraded detector (ITk). The thesis also presents the development of a tool for automatic visual inspection of the ITk pixel modules. This tool is based on a machine learning anomaly detection algorithm, and is designed to help the operations during production
Chariot, Alexandre. "Quelques applications de la programmation des processeurs graphiques à la simulation neuronale et à la vision par ordinateur." Phd thesis, Ecole des Ponts ParisTech, 2008. http://pastel.archives-ouvertes.fr/pastel-00005176.
Full textZhou, Rongyan. "Exploration of opportunities and challenges brought by Industry 4.0 to the global supply chains and the macroeconomy by integrating Artificial Intelligence and more traditional methods." Electronic Thesis or Diss., université Paris-Saclay, 2021. http://www.theses.fr/2021UPAST037.
Full textIndustry 4.0 is a significant shift and a tremendous challenge for every industrial segment, especially for the manufacturing industry that gave birth to the new industrial revolution. The research first uses literature analysis to sort out the literature, and focuses on the use of “core literature extension method” to enumerate the development direction and application status of different fields, which devotes to showing a leading role for theory and practice of industry 4.0. The research then explores the main trend of multi-tier supply in Industry 4.0 by combining machine learning and traditional methods. Next, the research investigates the relationship of industry 4.0 investment and employment to look into the inter-regional dependence of industry 4.0 so as to present a reasonable clustering based on different criteria and make suggestions and analysis of the global supply chain for enterprises and organizations. Furthermore, our analysis system takes a glance at the macroeconomy. The combination of natural language processing in machine learning to classify research topics and traditional literature review to investigate the multi-tier supply chain significantly improves the study's objectivity and lays a solid foundation for further research. Using complex networks and econometrics to analyze the global supply chain and macroeconomic issues enriches the research methodology at the macro and policy level. This research provides analysis and references to researchers, decision-makers, and companies for their strategic decision-making