Le projet scientifique dans lequel s’inscrit ma thèse a pour objectif l’élaboration d’algorithmes distribués et efficaces pour les réseaux informatiques. Ce projet vise une catégorie particulière des algorithmes distribués, dits auto-stabilisants. Il s’agit d’algorithmes ayant pour propriété de retrouver un comportement correct suite à une panne dans le réseau et ce, sans aucune intervention humaine. Le travail effectué en collaboration avec mes directeurs de thèse s’est concentré, plus précisément, autour des problèmes de couplage, de cliques et des paradigmes de publications-souscriptions dans ce domaine de l’informatique théorique. Dans un premier temps on a traité le problème du couplage maximal dans sa version anonyme, en fournissant un algorithme auto-stabilisant probabiliste et efficace. Ces travaux sont parus dans le journal PPL. De plus, on s’est intéressé au problème du couplage dans sa version maximum identifiée. Son travail améliore le dernier algorithme présent dans la littérature pour l’approximation de ce type de couplage au 2/3 de la solution optimale. Ces travaux sont parus dans une conférence internationale OPODIS. Par ailleurs, j'ai eu l’opportunité de collaborer en Allemagne avec Prof. Volker Turau au sein du groupe de télématique de l’Université technique de Hambourg. Le cadre de cette collaboration a été les algorithmes auto-stabilisants pour les paradigmes de publication-souscription. Cela a abouti à un algorithme efficace pour la version en canal de ce problème, introduisant la notion de raccourci pour le routage de messages dans ces paradigmes. Les résultats ont fait l’objet d’un Brief Announcement et d’un papier, publiés dans des conférences internationales, SSS et NetSyS. J'ai aussi bénéficié d’une collaboration avec Mr. Gerry Siegemund qui a été accueilli au laboratoire d’Informatique de l’École Polytechnique. Il a été question de trouver un algorithme efficace et auto-stabilisant pour la partition d’un réseau en cliques. Cette collaboration a eu pour résultat un algorithme pour le problème améliorant le dernier en date. Ce résultat est en cours de rédaction pour soumission à une conférence internationale. / The main focus of my thesis is the design of an efficient kind of distributed algorithms, known as: Self-stabilizing. These algorithms have the property to recover from faults in the environment they're executed in, and this without any human intervention. Recovering here, means converging toward a pre-defined, correct configuration. In this setting, I was mainly interested by the problems of matching in graphs, clique partitions and publication subscription paradigms. For the maximal version of the matching problem in anonymous graphs, we achieved a more efficient randomized, self-stabilizing algorithm. This work is published in a journal version in PPL. The maximum version of the same problem, but in an identified setting, led to the design of an efficient self-stabilizing algorithm that approximates the optimal solution up to the 2/3. This result was published at OPODIS. During a research visit at TUHH, Hamburg, Germany. Together with Pr. Volker Turau we tackled the problem of self-stabilizing publish/subscribe paradigms. This led to an algorithm introducing the new notion of short-cuts in this type of structures and was published under a brief announcement and a regular paper at SSS and NetSyS. In collaboration with Mr. Siegemund, then a visiting researcher at LIX, École Polytechnique, we worked on an efficient self-stabilizing algorithm for clique partitions. This work is still in progress and in preparation for an eventual publication.

Image representation plays an important role in image processing applications, which usually. contain a huge amount of data. An image is a two-dimensional array of points, and each point contains information (eg: color). A 1024 by 1024 pixel image occupies 1 mega byte of space in the main memory. In actual circumstances 2 to 3 mega bytes of space are needed to facilitate the various image processing tasks. Large amounts of secondary memory are also required to hold various data sets. In this thesis, two different operations on the quadtree are presented. There are, in general, two types of data compression techniques in image processing. One approach is based on elimination of redundant data from the original picture. Other techniques rely on higher levels of processing such as interpretations, generations, inductions and deduction procedures (1, 2). One of the popular techniques of data representation that has received a considerable amount of attention in recent years is the quadtree data structure. This has led to the development of various techniques for performing conversions and operations on the quadtree. Klinger and Dyer (3) provide a good bibliography of the history of quadtrees. Their paper reports experiments on the degree of compaction of picture representation which may be achieved with tree encoding. Their experiments show that tree encoding can produce memory savings. Pavlidis [15] reports on the approximation of pictures by quadtrees. Horowitz and Pavidis [16] show how to segment a picture using traversal of a quadtree. They segment the picture by polygonal boundaries. Tanimoto [17] discusses distortions which may occur in quadtrees for pictures. Tanimoto [18, p. 27] observes that quadtree representation is particularly convenient for scaling a picture by powers of two. Quadtrees are also useful in graphics and animation applications [19, 20] which are oriented toward construction of images from polygons and superpositiofis of images. Encoded pictures are useful for display, especially if encoding lends itself to processing.

In this thesis a heuristic method for factoring semiprimes by multiagent depth-limited search of PG2N graphs is presented. An analysis of PG2N graph connectivity is used to generate heuristics for multiagent search. Further analysis is presented including the requirements on choosing prime numbers to generate 'hard' semiprimes; the lack of connectivity in PG1N graphs; the counts of spanning trees in PG2N graphs; the upper bound of a PG2N graph diameter and a conjecture on the frequency distribution of prime numbers on Hamming distance. We further demonstrated the feasibility of the HD2 breadth first search of PG2N graphs for factoring small semiprimes. We presented the performance of different multiagent search heuristics in PG2N graphs showing that the heuristic of most connected seedpick outperforms least connected or random connected seedpick heuristics on small PG2N graphs of size N

In this thesis, available tools for call graph generation, processing and visualization are analyzed. Based on this analysis, a call-graph processing tool is designed. The tool is then implemented and tested on call graphs generated from various real-world programs, including the Linux Kernel.

No description available.

No description available.

No description available.

No description available.

Convolutional Neural Networks (CNNs) have dominated the task of imageclassification since 2012. Some key components of their success are that the underlying architecture integrates a set inductive biases such as translational invariance and the training computation can be significantly reduced by employing weight sharing. CNNs are powerful tools for generating new representations of images tailored to a particular task such as classification. However, because each image is passed through the network independent of other images, CNNs are not able to effectively aggregate information between examples. In this thesis, we explore the idea of using Graph Neural Networks (GNNs) in conjunction with CNNs to produce an architecture that has both the representational capacity of a CNN and the ability to aggregate information between examples. Graph Neural Networks apply the concept of convolutions directly on graphs. A result of this is that GNNs are able to learn from the connections between nodes. However, when working with image datasets, there is no obvious choice on how to construct a graph. There are certain heuristics such as ensuring homophily that have empirically been shown to increase the performance of GNNs. In this thesis, we apply different schemes of constructing a graph from image data for the downstream task of image classification and experiment with settings such as using multiple feature spaces and enforcing a bipartite graph structure. We also propose a model that allows for end to end training using CNNs and GNNs with proxies and attention that improves classification accuracy in comparison to a regular CNN. / Computer and Information Science

No description available.