Enterprises use live video streaming as a mean of communication. Streaming high-quality video to thousands of devices in a corporate network is not an easy task; the bandwidth requirements often exceed the network capacity. For that matter, Peer-To-Peer (P2P) networks have been proven beneficial, as peers can exchange content efficiently by utilizing the topology of the corporate network. However, such networks are dynamic and their topology might not always be known. In this project we propose ABD, a new dynamic graph representation learning approach, which aims to estimate the bandwidth capacity between peers in a corporate network. The architecture of ABDis adapted to the properties of corporate networks. The model is composed of an attention mechanism and a decoder. The attention mechanism produces node embeddings, while the decoder converts those embeddings into bandwidth predictions. The model aims to capture both the dynamicity and the structure of the dynamic network, using an advanced training process. The performance of ABD is tested with two dynamic graphs which were produced by real corporate networks. Our results show that ABD achieves better results when compared to existing state-of-the-art dynamic graph representation learning models. / Företag använder live video streaming för både intern och extern kommunikation. Strömmning av hög kvalitet video till tusentals tittare i ett företagsnätverk är inte enkelt eftersom bandbreddskraven ofta överstiger kapaciteten på nätverket. För att minska lasten på nätverket har Peer-to-Peer (P2P) nätverk visat sig vara en lösning. Här anpassar sig P2P nätverket efter företagsnätverkets struktur och kan därigenom utbyta video data på ett effektivt sätt. Anpassning till ett företagsnätverk är ett utmanande problem eftersom dom är dynamiska med förändring över tid och kännedom över topologin är inte alltid tillgänglig. I det här projektet föreslår vi en ny lösning, ABD, en dynamisk approach baserat på inlärning av grafrepresentationer. Vi försöker estimera den bandbreddskapacitet som finns mellan två peers eller tittare. Architekturen av ABD anpassar sig till egenskaperna av företagsnätverket. Själva modellen bakom ABD använder en koncentrationsmekanism och en avkodare. Attention mekanismen producerar node embeddings, medan avkodaren konverterar embeddings till estimeringar av bandbredden. Modellen fångar upp dynamiken och strukturen av nätverket med hjälp av en avancerad träningsprocess. Effektiviteten av ABD är testad på två dynamiska nätverksgrafer baserat på data från riktiga företagsnätverk. Enligt våra experiment har ABD bättre resultat när man jämför med andra state-of the-art modeller för inlärning av dynamisk grafrepresentation.

Neural networks trained using gradient-based optimization methods exhibit a surprising phenomenon known as mode connectivity, where two independently trained network weights are not isolated low loss minima in the parameter space. Instead, they can be connected by simple curves along which the loss remains low. In case of linear mode connectivity up to permutation, even linear interpolations of the trained weights incur low loss when networks that differ by permutation of their hidden neurons are considered equivalent. While some recent research suggest that this implies existence of a single near-convex loss basin to which the parameters converge, others have empirically shown distinct basins corresponding to different strategies to solve the task. In some settings, averaging multiple network weights naively, without explicitly accounting for permutation invariance still results in a network with improved generalization. In this thesis, linear mode connectivity among a set of neural networks independently trained on labelled datasets, both naively and upon reparameterization to account for permutation invariance is studied. Specifically, the effect of hidden layer width on the connectivity is empirically evaluated. The experiments are conducted on a two dimensional toy classification problem, and the insights are extended to deeper networks trained on handwritten digits and images. It is argued that accounting for permutation of hidden neurons either explicitly or implicitly is necessary for weight averaging to improve test performance. Furthermore, the results indicate that the training dynamics induced by the optimization plays a significant role, and large model width alone may not be a sufficient condition for linear model connectivity. / Neurala nätverk som tränats med gradientbaserade optimeringsmetoder uppvisar ett överraskande fenomen som kallas modeconnectivity, där två oberoende tränade nätverksvikter inte är isolerade lågförlustminima i parameterutrymmet. Istället kan de kopplas samman med enkla kurvor längs vilka förlusten förblir låg. I händelse av linjär mode-anslutning upp till permutation medför även linjära interpolationer av de tränade vikterna låga förluster när nätverk som skiljer sig åt genom permutation av deras dolda neuroner anses vara likvärdiga. Medan en del nyare undersökningar tyder på att detta innebär att det finns en enda nära-konvex förlustbassäng till vilken parametrarna konvergerar, har andra empiriskt visat distinkta bassänger som motsvarar olika strategier för att lösa uppgiften. I vissa inställningar resulterar ett naivt medelvärde av flera nätverksvikter, utan att uttryckligen ta hänsyn till permutationsinvarians, fortfarande i ett nätverk med förbättrad generalisering. I den här avhandlingen studeras linjärmodsanslutningar mellan en uppsättning neurala nätverk som är oberoende tränade på märkta datamängder, både naivt och vid omparameterisering för att ta hänsyn till permutationsinvarians. Specifikt utvärderas effekten av dold lagerbredd på anslutningen empiriskt. Experimenten utförs på ett tvådimensionellt leksaksklassificeringsproblem, och insikterna utökas till djupare nätverk som tränas på handskrivna siffror och bilder. Det hävdas att redogörelse för permutation av dolda neuroner antingen explicit eller implicit är nödvändigt för viktgenomsnitt för att förbättra testprestanda. Dessutom indikerar resultaten att träningsdynamiken som induceras av optimeringen spelar en betydande roll, och enbart stor modellbredd kanske inte är ett tillräckligt villkor för linjär modellanslutning.

Dans l'apprentissage par renforcement profond basé sur la valeur avec des mémoires de relecture, le paramètre de taille de lot joue un rôle crucial en déterminant le nombre de transitions échantillonnées pour chaque mise à jour de gradient. Étonnamment, malgré son importance, ce paramètre n'est généralement pas ajusté lors de la proposition de nouveaux algorithmes. Dans ce travail, nous menons une vaste étude empirique qui suggère que la réduction de la taille des lots peut entraîner un certain nombre de gains de performances significatifs ; ceci est surprenant et contraire à la pratique courante consistant à utiliser de plus grandes tailles de lots pour améliorer la formation du réseau neuronal. Ce résultat inattendu défie la sagesse conventionnelle et appelle à une compréhension plus approfondie des gains de performances observés associés à des tailles de lots plus petites. Pour faire la lumière sur les facteurs sous-jacents, nous complétons nos résultats expérimentaux par une série d'analyses empiriques. Ces analyses approfondissent divers aspects du processus d'apprentissage, tels que l'analyse de la dynamique d'optimisation du réseau, la vitesse de convergence, la stabilité et les capacités d'exploration. Le chapitre 1 présente les concepts nécessaires pour comprendre le travail présenté, notamment des aperçus de l'Apprentissage Profond (Deep Learning) et de l'Apprentissage par Renforcement (Reinforcement Learning). Le chapitre 2 contient une description détaillée de nos contributions visant à comprendre les gains de performance observés associés à des tailles de lots plus petites lors de l'utilisation d'algorithmes d'apprentissage par renforcement profond basés sur la valeur. À la fin, des conclusions tirées de ce travail sont fournies, incluant des suggestions pour des travaux futurs. Le chapitre 3 aborde ce travail dans le contexte plus large de la recherche en apprentissage par renforcement. / In value-based deep reinforcement learning with replay memories, the batch size parameter plays a crucial role by determining the number of transitions sampled for each gradient update. Surprisingly, despite its importance, this parameter is typically not adjusted when proposing new algorithms. In this work, we conduct a broad empirical study that suggests {\em reducing} the batch size can result in a number of significant performance gains; this is surprising and contrary to the prevailing practice of using larger batch sizes to enhance neural network training. This unexpected result challenges the conventional wisdom and calls for a deeper understanding of the observed performance gains associated with smaller batch sizes. To shed light on the underlying factors, we complement our experimental findings with a series of empirical analyses such as analysis of network optimization dynamics, convergence speed, stability, and exploration capabilities. Chapter 1 introduces concepts necessary to understand the work presented, including overviews of Deep Learning and Reinforcement Learning. Chapter 2 contains a detailed description of our contributions towards understanding the observed performance gains associated with smaller batch sizes when using value based deep reinforcement learning algorithms. At the end, some conclusions drawn from this work are provided, including some exciting suggestion as future work. Chapter 3 talks about this work in the broader context of reinforcement learning research.

The last decade has brought with it a wave of innovative technology, shifting the channels through which creative content is created, consumed, and categorized. And, as our interactions with creative multimedia content shift towards online platforms, the sheer quantity of content on these platforms has necessitated the integration of algorithmic guidance in the discovery of these spaces. In this way, the recommendation algorithms that guide users' interactions with various art forms have been cast into the role of gatekeepers and begun to play an increasingly influential role in shaping the creation of artistic content. The work laid out in the following chapters fuses three major areas of research: graph representation learning, music information retrieval, and fairness as applied to the task of music recommendation. In recent years, graph neural networks (GNNs), a powerful new architecture which enables deep learning approaches to be applied to graph or network structures, have proven incredibly influential in the music recommendation domain. In tandem with the striking performance gains that GNNs are able to achieve, many of these systems, have been shown to be strongly influenced by the degree, or number of outgoing edges, of individual nodes. More concretely, recent works have uncovered disparities in the qualities of representations learned by state of the art GNNs between nodes which are strongly and weakly connected. Translating these findings to the sphere of recommender systems, where nodes and edges are used to represent the interactions between users and various items, these disparities in representation that are contingent upon a node's connectivity can be seen as a form of popularity bias. And, indeed, within the broader recommendation community, popularity bias has long been considered an open problem, in which recommender systems begin to favor mainstream content over, potentially more relevant, but niche or novel items. If left unchecked these algorithmic nudged towards previously popular content can create, intensify, and enforce negative cycles that perpetuate disparities in representation on both the user and the creator ends of the content consumption pipeline. Particularly in the recommendation of creative (e.g. musical) content, the downstream effects in these disparities of visibility can have genuine economic consequences for artists from under-represented communities. Thus, the problem of popularity bias is something that must be addressed from both a technical and societal perspective. And, as the influence of recommender systems continues to spread, the effects of this phenomenon only become more spurious, as they begin to have critical downstream effects that shape the larger ecosystems in which art is created. Thus, the broad focus of thesis is the mitigation of popularity bias in music recommendation. In order to tailor our exploration of this issue to the graph domain, we begin by formalizing the relationship between degree fairness and popularity bias. In doing so, we concretely define the notion of popularity, grounding it in the structural principles of an interaction network, and enabling us to design objectives that can mitigate the effects of popularity on representation learning. In our first work, we focus on understanding the effects of sampling on degree fairness in uni-partite graphs. The purpose of this work is to lay the foundation for the graph neural network model which will underlie our music recommender system. We then build off this first work by extending the initial fairness framework to be compatible with bi-partite graphs and applying it to the music domain. The motivation of this work is rooted in the notion of discovery, or the idea that users engage with algorithmic curation in order to find content that is both novel and relevant to their artistic tastes. We present the intrinsic relationship between discovery objectives and the presence of popularity bias, explaining that the presence of popularity bias can blind a system to the musical qualities that underpin the underlying needs of music listening. As we will explain in later sections, one of the key elements of this work is our ability to ground our fairness notion in the musical domain. Thus, we propose a domain-aware, individual fairness-based approach which addresses popularity bias in graph neural network (GNNs) based recommender systems. In order to facilitate this domain awareness, we perform extensive dataset augmentation, taking two state of the art music recommendation datasets and augmenting them with rich multi-modal node-level features. Finally, we ground our evaluation in the cold start setting, showing the importance of inductive methodologies in the music space. / La dernière décennie a apporté avec elle une vague de technologies innovantes, modifiant la manière dont le contenu créatif est créé, consommé et catégorisé. Et, à mesure que nos interactions avec les contenus multimédias créatifs se déplacent vers les plateformes en ligne, la quantité de contenu sur ces plateformes a nécessité l’intégration d’un guidage algorithmique dans la découverte de ces espaces. De cette façon, les algorithmes de recommandation qui guident les interactions des utilisateurs avec diverses formes d’art ont été jetés dans le rôle de gardiens et ont commencé à jouer un rôle de plus en plus influent dans l’élaboration de la création de contenu artistique. Le travail présenté dans les chapitres suivants fusionne trois grands domaines de recherche : l’apprentissage de la représentation graphique, la recherche d’informations musicales et l’équité appliquée à la tâche de recommandation musicale. Alors que l’influence des systèmes de recommandation continue de s’étendre et de s’intensifier, il est crucial de prendre en compte les effets en aval que les choix de conception peuvent avoir sur l’écosystème plus large de la création artistique. Ces dernières années, l’intégration des réseaux sociaux dans la tâche de recommandation musicale a donné naissance aux réseaux neuronaux de graphes (GNN), une nouvelle architecture capable de faire des prédictions sur les structures de graphes. Parallèlement aux gains miraculeux que les GNN sont capables de réaliser, bon nombre de ces systèmes peuvent également être la proie de biais de popularité, les forçant à privilégier le contenu grand public par rapport à des éléments potentiellement plus pertinents, mais de niche ou nouveaux. S’il n’est pas maîtrisé, ce cycle négatif peut perpétuer les disparités de représentation entre la musique d’artistes, de genres ou de populations minoritaires. Et, ce faisant, les disparités dans la visibilité des éléments peuvent entraîner des problèmes à la fois du point de vue des performances et de la société. L’objectif de la thèse est l’atténuation du biais de popularité. Premièrement, le travail formalise les liens entre l’équité individuelle et la présence d’un biais de popularité parmi les contenus créatifs. Ensuite, nous étendons un cadre d’équité individuelle, en l’appliquant au domaine de la recommandation musicale. Le coeur de cette thèse s’articule autour de la proposition d’une approche basée sur l’équité individuelle et sensible au domaine qui traite le biais de popularité dans les systèmes de recommandation basés sur les réseaux de 5 neurones graphiques (GNN). L’un des éléments clés de ce travail est notre capacité à ancrer notre notion d’équité dans le domaine musical. Afin de faciliter cette prise de conscience du domaine, nous effectuons une augmentation étendue des ensembles de données, en prenant deux ensembles de données de recommandation musicale à la pointe de la technologie et en les augmentant avec de riches fonctionnalités multimodales au niveau des noeuds. Enfin, nous fondons notre évaluation sur le démarrage à froid, montrant l’importance des méthodologies inductives dans l’espace musical.

Representation Learning is ubiquitous in state-of-the-art machine learning workflow, including data exploration/visualization, data preprocessing, data model learning, and model interpretations. However, the majority of the newly proposed Representation Learning methods are more suitable for problems with a large amount of data. Applying these methods to problems with a limited amount of data may lead to unsatisfactory performance. Therefore, there is a need for developing Representation Learning methods which are tailored for problems with ``small data", such as, clinical and biomedical data analytics. In this dissertation, we describe our studies of tackling the challenging clinical and biomedical data analytics problem from four perspectives: data preprocessing, temporal data representation learning, output representation learning, and joint input-output representation learning. Data scaling is an important component in data preprocessing. The objective in data scaling is to scale/transform the raw features into reasonable ranges such that each feature of an instance will be equally exploited by the machine learning model. For example, in a credit flaw detection task, a machine learning model may utilize a person's credit score and annual income as features, but because the ranges of these two features are different, a machine learning model may consider one more heavily than another. In this dissertation, I thoroughly introduce the problem in data scaling and describe an approach for data scaling which can intrinsically handle the outlier problem and lead to better model prediction performance. Learning new representations for data in the unstandardized form is a common task in data analytics and data science applications. Usually, data come in a tubular form, namely, the data is represented by a table in which each row is a feature (row) vector of an instance. However, it is also common that the data are not in this form; for example, texts, images, and video/audio records. In this dissertation, I describe the challenge of analyzing imperfect multivariate time series data in healthcare and biomedical research and show that the proposed method can learn a powerful representation to encounter various imperfections and lead to an improvement of prediction performance. Learning output representations is a new aspect of Representation Learning, and its applications have shown promising results in complex tasks, including computer vision and recommendation systems. The main objective of an output representation algorithm is to explore the relationship among the target variables, such that a prediction model can efficiently exploit the similarities and potentially improve prediction performance. In this dissertation, I describe a learning framework which incorporates output representation learning to time-to-event estimation. Particularly, the approach learns the model parameters and time vectors simultaneously. Experimental results do not only show the effectiveness of this approach but also show the interpretability of this approach from the visualizations of the time vectors in 2-D space. Learning the input (feature) representation, output representation, and predictive modeling are closely related to each other. Therefore, it is a very natural extension of the state-of-the-art by considering them together in a joint framework. In this dissertation, I describe a large-margin ranking-based learning framework for time-to-event estimation with joint input embedding learning, output embedding learning, and model parameter learning. In the framework, I cast the functional learning problem to a kernel learning problem, and by adopting the theories in Multiple Kernel Learning, I propose an efficient optimization algorithm. Empirical results also show its effectiveness on several benchmark datasets. / Computer and Information Science

Knowledge graphs (KGs) store relational information in a flexible triplet schema and have become ubiquitous for information storage in domains such as web search, e-commerce, social networks, and biology. Retrieval of information from KGs is generally achieved through logical reasoning, but this process can be computationally expensive and has limited performance due to the large size and complexity of relationships within the KGs. Furthermore, to extend the usage of KGs to non-expert users, retrieval over them cannot solely rely on logical reasoning but also needs to consider text-based search. This creates a need for multi-modal representations that capture both the semantic and structural features from the KGs. The primary objective of the proposed work is to extend the accessibility of KGs to non-expert users/institutions by enabling them to utilize non-technical textual queries to search over the vast amount of information stored in KGs. To achieve this objective, the research aims to solve four limitations: (i) develop a framework for logical reasoning over KGs that can learn representations to capture hierarchical dependencies between entities, (ii) design an architecture that can effectively learn the logic flow of queries from natural language text, (iii) create a multi-modal architecture that can capture inherent semantic and structural features from the entities and KGs, respectively, and (iv) introduce a novel hyperbolic learning framework to enable the scalability of hyperbolic neural networks over large graphs using meta-learning. The proposed work is distinct from current research because it models the logical flow of textual queries in hyperbolic space and uses it to perform complex reasoning over large KGs. The models developed in this work are evaluated on both the standard research setting of logical reasoning, as well as, real-world scenarios of query matching and search, specifically, in the e-commerce domain. In summary, the proposed work aims to extend the accessibility of KGs to non-expert users by enabling them to use non-technical textual queries to search vast amounts of information stored in KGs. To achieve this objective, the work proposes the use of multi-modal representations that capture both semantic and structural features from the KGs, and a novel hyperbolic learning framework to enable scalability of hyperbolic neural networks over large graphs. The work also models the logical flow of textual queries in hyperbolic space to perform complex reasoning over large KGs. The models developed in this work are evaluated on both the standard research setting of logical reasoning and real-world scenarios in the e-commerce domain. / Doctor of Philosophy / Knowledge graphs (KGs) are databases that store information in a way that allows computers to easily identify relationships between different pieces of data. They are widely used in domains such as web search, e-commerce, social networks, and biology. However, retrieving information from KGs can be computationally expensive, and relying solely on logical reasoning can limit their accessibility to non-expert users. This is where the proposed work comes in. The primary objective is to make KGs more accessible to non-experts by enabling them to use natural language queries to search the vast amounts of information stored in KGs. To achieve this objective, the research aims to address four limitations. Firstly, a framework for logical reasoning over KGs that can learn representations to capture hierarchical dependencies between entities is developed. Secondly, an architecture is designed that can effectively learn the logic flow of queries from natural language text. Thirdly, a multi-modal architecture is created that can capture inherent semantic and structural features from the entities and KGs, respectively. Finally, a novel hyperbolic learning framework is introduced to enable the scalability of hyperbolic neural networks over large graphs using meta-learning. The proposed work is unique because it models the logical flow of textual queries in hyperbolic space and uses it to perform complex reasoning over large KGs. The models developed in this work are evaluated on both the standard research setting of logical reasoning, as well as, real-world scenarios of query matching and search, specifically, in the e-commerce domain. In summary, the proposed work aims to make KGs more accessible to non-experts by enabling them to use natural language queries to search vast amounts of information stored in KGs. To achieve this objective, the work proposes the use of multi-modal representations that capture both semantic and structural features from the KGs, and a novel hyperbolic learning framework to enable scalability of hyperbolic neural networks over large graphs. The work also models the logical flow of textual queries in hyperbolic space to perform complex reasoning over large KGs. The results of this work have significant implications for the field of information retrieval, as it provides a more efficient and accessible way to retrieve information from KGs. Additionally, the multi-modal approach taken in this work has potential applications in other areas of machine learning, such as image recognition and natural language processing. The work also contributes to the development of hyperbolic geometry as a tool for modeling complex networks, which has implications for fields such as network science and social network analysis. Overall, this work represents an important step towards making the vast amounts of information stored in KGs more accessible and useful to a wider audience.

Mobile cellular networks are widely integrated in today’s infrastructure. These networks are constantly evolving and continuously expanding, especially with the introduction of fifth-generation (5G). It is important to ensure the effectiveness of these expansions.Mobile networks consist of a set of radio nodes that are distributed in a geographicalregion to provide connectivity services. Each radio node is served by a set of cells. Thehandover relations between cells is determined by Software features such as AutomaticNeighbor Relations (ANR). The handover relations, also refereed as edges, betweenradio nodes in the mobile network graph are created through historical interactions between User Equipment (UE) and radio nodes. The method has the limitation of not being able to set the edges before the physical hardware is integrated. In this work, we usegraph-based deep learning methods to determine mobility relations (edges), trained onradio node configuration data and a set of reliable relations of ANR in stable networks.The report focuses on measuring the accuracy and precision of different graph baseddeep learning approaches applied to real-world mobile networks. The report considers four models. Our comprehensive experiments on Telecom datasets obtained fromoperational Telecom Networks demonstrate that graph neural network model and multilayer perceptron trained with Binary Cross Entropy (BCE) loss outperform all othermodels. The four models evaluation showed that considering graph structure improveresults. Additionally, the model investigates the use of heuristics to reduce the trainingtime based on distance between radio node to eliminate irrelevant cases. The use ofthese heuristics improved precision and accuracy.

<p>Research on graph representation learning (GRL) has made major strides over the past decade, with widespread applications in domains such as e-commerce, personalization, fraud & abuse, life sciences, and social network analysis. Despite its widespread success, fundamental questions on practices employed in modern day GRL have remained unanswered. Unraveling and advancing two such fundamental questions on the practices in modern day GRL forms the overarching theme of my thesis.</p> <p>The first part of my thesis deals with the mathematical foundations of GRL. GRL is used to solve tasks such as node classification, link prediction, clustering, graph classification, and so on, albeit with seemingly different frameworks (e.g. Graph neural networks for node/graph classification, (implicit) matrix factorization for link prediction/ clustering, etc.). The existence of very distinct frameworks for different graph tasks has puzzled researchers and practitioners alike. In my thesis, using group theory, I provide a theoretical blueprint that connects these seemingly different frameworks, bridging methods like matrix factorization and graph neural networks. With this renewed understanding, I then provide guidelines to better realize the full capabilities of these methods in a multitude of tasks.</p> <p>The second part of my thesis deals with cases where modeling real-world objects as a graph is an oversimplified description of the underlying data. Specifically, I look at two such objects (i) modeling hypergraphs (where edges encompass two or more vertices) and (ii) using GRL for predicting protein properties. Towards (i) hypergraphs, I develop a hypergraph neural network which takes advantage of the inherent sparsity of real world hypergraphs, without unduly sacrificing on its ability to distinguish non isomorphic hypergraphs. The designed hypergraph neural network is then leveraged to learn expressive representations of hyperedges for two tasks, namely hyperedge classification and hyperedge expansion. Experiments show that using our network results in improved performance over the current approach of converting the hypergraph into a dyadic graph and using (dyadic) GRL frameworks. Towards (ii) proteins, I introduce the concept of conditional invariances and leverage it to model the inherent flexibility present in proteins. Using conditional invariances, I provide a new framework for GRL which can capture protein-dependent conformations and ensures that all viable conformers of a protein obtain the same representation. Experiments show that endowing existing GRL models with my framework shows noticeable improvements on multiple different protein datasets and tasks.</p>

Cette thèse aborde le problème de la recherche d'images à grande échelle. Pour aborder la recherche d'images à grande échelle, il est nécessaire de coder des images avec des représentations compactes qui peuvent être efficacement utilisées pour comparer des images de manière significative. L'obtention d'une telle représentation compacte peut se faire soit en comprimant des représentations efficaces de grande dimension, soit en apprenant des représentations compactes de bout en bout. Le travail de cette thèse explore et avance dans ces deux directions. Dans notre première contribution, nous étendons les approches de quantification vectorielle structurée telles que la quantification de produit en proposant une représentation somme pondérée de codewords. Nous testons et vérifions les avantages de notre approche pour la recherche approximative du plus proche voisin sur les caractéristiques d'image locales et globales, ce qui est un moyen important d'aborder la recherche d'images à grande échelle. L'apprentissage de la représentation compacte pour la recherche d'images a récemment attiré beaucoup d'attention avec diverses approches basées sur le hachage profond proposées. Dans de telles approches, les réseaux de neurones convolutifs profonds apprennent à coder des images en codes binaires compacts. Dans cette thèse, nous proposons une approche d'apprentissage supervisé profond pour la représentation binaire structurée qui rappelle une approche de quantification vectorielle structurée telle que PQ. Notre approche bénéficie de la recherche asymétrique par rapport aux approches de hachage profond et apporte une nette amélioration de la précision de la recherche au même débit binaire. L'index inversé est une autre partie importante du système de recherche à grande échelle en dehors de la représentation compacte. À cette fin, nous étendons nos idées pour l'apprentissage de la représentation compacte supervisée pour la construction d'index inversés. Dans ce travail, nous abordons l'indexation inversée avec un apprentissage approfondi supervisé et essayons d'unifier l'apprentissage de l'indice inversé et de la représentation compacte. Nous évaluons minutieusement toutes les méthodes proposées sur divers ensembles de données accessibles au public. Nos méthodes surpassent ou sont compétitives avec l'état de l'art. / This thesis addresses the problem of large-scale image search. To tackle image search at large scale, it is required to encode images with compact representations which can be efficiently employed to compare images meaningfully. Obtaining such compact representation can be done either by compressing effective high dimensional representations or by learning compact representations in an end-to-end manner. The work in this thesis explores and advances in both of these directions. In our first contribution, we extend structured vector quantization approaches such as Product Quantization by proposing a weighted codeword sum representation. We test and verify the benefits of our approach for approximate nearest neighbor search on local and global image features which is an important way to approach large scale image search. Learning compact representation for image search recently got a lot of attention with various deep hashing based approaches being proposed. In such approaches, deep convolutional neural networks are learned to encode images into compact binary codes. In this thesis we propose a deep supervised learning approach for structured binary representation which is a reminiscent of structured vector quantization approaches such as PQ. Our approach benefits from asymmetric search over deep hashing approaches and gives a clear improvement for search accuracy at the same bit-rate. Inverted index is another important part of large scale search system apart from the compact representation. To this end, we extend our ideas for supervised compact representation learning for building inverted indexes. In this work we approach inverted indexing with supervised deep learning and make an attempt to unify the learning of inverted index and compact representation. We thoroughly evaluate all the proposed methods on various publicly available datasets. Our methods either outperform, or are competitive with the state-of-the-art.

Différentes disciplines des sciences humaines telles la philologie ou la paléographie font face à des tâches complexes et fastidieuses pour l'examen des sources de données. La proposition d'approches computationnelles en humanités permet d'adresser les problématiques rencontrées telles que la lecture, l'analyse et l'archivage de façon systématique. Les modèles conceptuels élaborés reposent sur des algorithmes et ces derniers donnent lieu à des implémentations informatiques qui automatisent ces tâches fastidieuses. La première partie de la thèse vise, d'une part, à établir la structuration thématique d'un corpus, en construisant des espaces sémantiques de grande dimension. D'autre part, elle vise au suivi dynamique des thématiques qui constitue un réel défi scientifique, notamment en raison du passage à l'échelle. La seconde partie de la thèse traite de manière holistique la page d'un document numérisé sans aucune intervention préalable. Le but est d'apprendre automatiquement des représentations du trait de l'écriture ou du tracé d'un certain script par rapport au tracé d'un autre script. Il faut dans ce cadre tenir compte de l'environnement où se trouve le tracé : image, artefact, bruits dus à la détérioration de la qualité du papier, etc. Notre approche propose un empilement de réseaux de neurones auto-encodeurs afin de fournir une représentation alternative des données reçues en entrée. / Different disciplines in the humanities, such as philology or palaeography, face complex and time-consuming tasks whenever it comes to examining the data sources. The introduction of computational approaches in humanities makes it possible to address issues such as semantic analysis and systematic archiving. The conceptual models developed are based on algorithms that are later hard coded in order to automate these tedious tasks. In the first part of the thesis we propose a novel method to build a semantic space based on topics modeling. In the second part and in order to classify historical documents according to their script. We propose a novel representation learning method based on stacking convolutional auto-encoder. The goal is to automatically learn plot representations of the script or the written language.