<p dir="ltr">Generative Image-to-Image Translation (I2IT) involves transforming an input image from one domain to another. Typically, this transformation retains the content in the input image while adjusting the domain-dependent style elements. Generative I2IT finds utility in a wide range of applications, yet its effectiveness hinges on adaptations to the unique characteristics of the data at hand. This dissertation pushes the boundaries of I2IT by applying it to stain-related problems in computational pathology. Particularly, the main contributions span two major applications of stain translation: H&E-to-H&E and H&E-to-IHC, each with its unique requirements and challenges. More specifically, the first contribution addresses the generalization challenge posed by the high variability in H&E stain appearances to any task-specific machine learning models. To this end, the Generative Stain Augmentation Network (G-SAN) is introduced to augment the training images in any downstream task with random and diverse H&E stain appearances. Experimental results demonstrate G-SAN’s ability to enhance model generalization across stain variations in downstream tasks. The second key contribution in this dissertation focuses on H&E-to-IHC stain translation. The major challenge in learning accurate H&E-to-IHC stain translation is the frequent and sometimes severe inconsistencies in the groundtruth H&E-IHC image pairs. To make training more robust to these inconsistencies, a novel contrastive learning based loss, named the Adaptive Supervised PatchNCE (ASP) loss is presented. Experimental results suggest that the proposed ASP-based framework outperforms the state-of-the-art in H&E-to-IHC stain translation by significant margins. Additionally, a new dataset for H&E-to-IHC translation – the Multi-IHC Stain Translation (MIST) dataset, is released to the public, featuring paired images from H&E to four different IHC stains. For future directions of generative I2IT in stain translation problems, a proof-of-concept study of applying the latest diffusion model based I2IT methods to the problem of virtual H&E staining is presented.</p>

Les formes, les surfaces, les événements et les objets (vivants et non vivants) constituent le monde. L'intelligence des agents naturels, tels que les humains, va au-delà de la simple reconnaissance de formes. Nous excellons à construire des représentations et à distiller des connaissances pour comprendre et déduire la structure du monde. Spécifiquement, le développement de telles capacités de raisonnement peut se produire même avec une supervision limitée. D'autre part, malgré son développement phénoménal, les succès majeurs de l'apprentissage automatique, en particulier des modèles d'apprentissage profond, se situent principalement dans les tâches qui ont accès à de grands ensembles de données annotées. Dans cette thèse, nous proposons de nouvelles solutions pour aider à combler cette lacune en permettant aux modèles d'apprentissage automatique d'apprendre la structure et de permettre un raisonnement efficace en présence de tâches faiblement supervisés. Le thème récurrent de la thèse tente de s'articuler autour de la question « Comment un système perceptif peut-il apprendre à organiser des informations sensorielles en connaissances utiles sous une supervision limitée ? » Et il aborde les thèmes de la géométrie, de la composition et des associations dans quatre articles distincts avec des applications à la vision par ordinateur (CV) et à l'apprentissage par renforcement (RL). Notre première contribution ---Pix2Shape---présente une approche basée sur l'analyse par synthèse pour la perception. Pix2Shape exploite des modèles génératifs probabilistes pour apprendre des représentations 3D à partir d'images 2D uniques. Le formalisme qui en résulte nous offre une nouvelle façon de distiller l'information d'une scène ainsi qu'une représentation puissantes des images. Nous y parvenons en augmentant l'apprentissage profond non supervisé avec des biais inductifs basés sur la physique pour décomposer la structure causale des images en géométrie, orientation, pose, réflectance et éclairage. Notre deuxième contribution ---MILe--- aborde les problèmes d'ambiguïté dans les ensembles de données à label unique tels que ImageNet. Il est souvent inapproprié de décrire une image avec un seul label lorsqu'il est composé de plus d'un objet proéminent. Nous montrons que l'intégration d'idées issues de la littérature linguistique cognitive et l'imposition de biais inductifs appropriés aident à distiller de multiples descriptions possibles à l'aide d'ensembles de données aussi faiblement étiquetés. Ensuite, nous passons au paradigme d'apprentissage par renforcement, et considérons un agent interagissant avec son environnement sans signal de récompense. Notre troisième contribution ---HaC--- est une approche non supervisée basée sur la curiosité pour apprendre les associations entre les modalités visuelles et tactiles. Cela aide l'agent à explorer l'environnement de manière autonome et à utiliser davantage ses connaissances pour s'adapter aux tâches en aval. La supervision dense des récompenses n'est pas toujours disponible (ou n'est pas facile à concevoir), dans de tels cas, une exploration efficace est utile pour générer un comportement significatif de manière auto-supervisée. Pour notre contribution finale, nous abordons l'information limitée contenue dans les représentations obtenues par des agents RL non supervisés. Ceci peut avoir un effet néfaste sur la performance des agents lorsque leur perception est basée sur des images de haute dimension. Notre approche a base de modèles combine l'exploration et la planification sans récompense pour affiner efficacement les modèles pré-formés non supervisés, obtenant des résultats comparables à un agent entraîné spécifiquement sur ces tâches. Il s'agit d'une étape vers la création d'agents capables de généraliser rapidement à plusieurs tâches en utilisant uniquement des images comme perception. / Shapes, surfaces, events, and objects (living and non-living) constitute the world. The intelligence of natural agents, such as humans is beyond pattern recognition. We excel at building representations and distilling knowledge to understand and infer the structure of the world. Critically, the development of such reasoning capabilities can occur even with limited supervision. On the other hand, despite its phenomenal development, the major successes of machine learning, in particular, deep learning models are primarily in tasks that have access to large annotated datasets. In this dissertation, we propose novel solutions to help address this gap by enabling machine learning models to learn the structure and enable effective reasoning in the presence of weakly supervised settings. The recurring theme of the thesis tries to revolve around the question of "How can a perceptual system learn to organize sensory information into useful knowledge under limited supervision?" And it discusses the themes of geometry, compositions, and associations in four separate articles with applications to computer vision (CV) and reinforcement learning (RL). Our first contribution ---Pix2Shape---presents an analysis-by-synthesis based approach(also referred to as inverse graphics) for perception. Pix2Shape leverages probabilistic generative models to learn 3D-aware representations from single 2D images. The resulting formalism allows us to perform a novel view synthesis of a scene and produce powerful representations of images. We achieve this by augmenting unsupervised learning with physically based inductive biases to decompose a scene structure into geometry, pose, reflectance and lighting. Our Second contribution ---MILe--- addresses the ambiguity issues in single-labeled datasets such as ImageNet. It is often inappropriate to describe an image with a single label when it is composed of more than one prominent object. We show that integrating ideas from Cognitive linguistic literature and imposing appropriate inductive biases helps in distilling multiple possible descriptions using such weakly labeled datasets. Next, moving into the RL setting, we consider an agent interacting with its environment without a reward signal. Our third Contribution ---HaC--- is a curiosity based unsupervised approach to learning associations between visual and tactile modalities. This aids the agent to explore the environment in an analogous self-guided fashion and further use this knowledge to adapt to downstream tasks. In the absence of reward supervision, intrinsic movitivation is useful to generate meaningful behavior in a self-supervised manner. In our final contribution, we address the representation learning bottleneck in unsupervised RL agents that has detrimental effect on the performance on high-dimensional pixel based inputs. Our model-based approach combines reward-free exploration and planning to efficiently fine-tune unsupervised pre-trained models, achieving comparable results to task-specific baselines. This is a step towards building agents that can generalize quickly on more than a single task using image inputs alone.

This thesis aims to investigate a fault detection and diagnosis system for automotive cameras using unsupervised learning. 1) Can a front-looking wide-angle camera image dataset be created using Hardware-in-Loop (HIL) simulations? 2) Can an Adversarial Autoencoder (AAE) based unsupervised camera fault detection and diagnosis method be crafted for SPA2 Vehicle Control Unit (VCU) using an image dataset created using Hardware-inLoop? 3) Does using AAE surpass the performance of using Variational Autoencoder (VAE) for the unsupervised automotive camera fault diagnosis model? In the field of camera fault studies, automotive cameras stand out for its complex operational context, particularly in Advanced Driver-Assistance Systems (ADAS) applications. The literature review finds a notable gap in comprehensive image datasets addressing the image artefact spectrum of ADAS-equipped automotive cameras under real-world driving conditions. In this study, normal and fault scenarios for automotive cameras are defined leveraging published and company studies and a fault diagnosis model using unsupervised learning is proposed and examined. The types of image faults defined and included are Lens Flare, Gaussian Noise and Dead Pixels. Along with normal driving images, a balanced fault-injected image dataset is collected using real-time sensor simulation under driving scenario with industrially-recognised HIL setup. An AAE-based unsupervised automotive camera fault diagnosis system using VGG16 as encoder-decoder structure is proposed and experiments on its performance are conducted on both the selfcollected dataset and fault-injected KITTI raw images. For non-processed KITTI dataset, morphological operations are examined and are employed as preprocessing. The performance of the system is discussed in comparison to supervised and unsupervised image partition methods in related works. The research found that the AAE method outperforms popular VAE method, using VGG16 as encoder-decoder structure significantly using 3-layer Convolutional Neural Network (CNN) and ResNet18 and morphological preprocessings significantly ameliorate system performance. The best performing VGG16- AAE model achieves 62.7% accuracy to diagnosis on own dataset, and 86.4% accuracy on double-erosion-processed fault-injected KITTI dataset. In conclusion, this study introduced a novel scheme for collecting automotive sensor data using Hardware-in-Loop, utilised preprocessing techniques that enhance image partitioning and examined the application of unsupervised models for diagnosing faults in automotive cameras. / Denna avhandling syftar till att undersöka ett felupptäcknings- och diagnossystem för bilkameror med hjälp av oövervakad inlärning. De huvudsakliga forskningsfrågorna är om en bilduppsättning från en frontmonterad vidvinkelkamera kan skapas med hjälp av Hardware-in-Loop (HIL)-simulationer, om en Adversarial Autoencoder (AAE)-baserad metod för oövervakad felupptäckt och diagnos för SPA2 Vehicle Control Unit (VCU) kan utformas med en bilduppsättning skapad med Hardware-in-Loop, och om användningen av AAE skulle överträffa prestandan av att använda Variational Autoencoder (VAE) för den oövervakade modellen för felanalys i bilkameror. Befintliga studier om felanalys fokuserar på roterande maskiner, luftbehandlingsenheter och järnvägsfordon. Få studier undersöker definitionen av feltyper i bilkameror och klassificerar normala och felaktiga bilddata från kameror i kommersiella passagerarfordon. I denna studie definieras normala och felaktiga scenarier för bilkameror och en modell för felanalys med oövervakad inlärning föreslås och undersöks. De typer av bildfel som definieras är Lens Flare, Gaussiskt brus och Döda pixlar. Tillsammans med normala bilder samlas en balanserad uppsättning felinjicerade bilder in med hjälp av realtidssensor-simulering under körscenarier med industriellt erkänd HIL-uppsättning. Ett AAE-baserat system för oövervakad felanalys i bilkameror med VGG16 som kodaredekoderstruktur föreslås och experiment på dess prestanda genomförs både på den självinsamlade uppsättningen och felinjicerade KITTI-raw-bilder. För icke-behandlade KITTI-uppsättningar undersöks morfologiska operationer och används som förbehandling. Systemets prestanda diskuteras i jämförelse med övervakade och oövervakade bildpartitioneringsmetoder i relaterade arbeten. Forskningen fann att AAE-metoden överträffar den populära VAEmetoden, genom att använda VGG16 som kodare-dekoderstruktur signifikant med ett 3-lagers konvolutionellt neuralt nätverk (CNN) och ResNet18 och morfologiska förbehandlingar förbättrar systemets prestanda avsevärt. Den bäst presterande VGG16-AAE-modellen uppnår 62,7 % noggrannhet för diagnos på egen uppsättning, och 86,4 % noggrannhet på dubbelerosionsbehandlad felinjicerad KITTI-uppsättning. Sammanfattningsvis introducerade denna studie ett nytt system för insamling av data från bilsensorer med Hardware-in-Loop, utnyttjade förbehandlingstekniker som förbättrar bildpartitionering och undersökte tillämpningen av oövervakade modeller för att diagnostisera fel i bilkameror.

A challenge to release autonomous vehicles to public roads is safety verification of the developed features. Safety test driving of vehicles is not practically feasible as the acceptance criterion is driving at least 2.1 billion kilometers [1]. An alternative to this distance-based testing is the scenario-based approach, where the intelligent vehicles are exposed to known scenarios. Identification of such scenarios from the driving data is crucial for this validation. The aim of this thesis is to investigate the possibility of unsupervised identification of driving scenarios from the driving data. The task is performed in two major parts. The first is the segmentation of the time series driving data by detecting changepoints, followed by the clustering of the previously obtained segments. Time-series segmentation is approached using a Deep Learning method, while the second task is performed using time series clustering. The work also includes a visual approach for validating the time-series segmentation, followed by a quantitative measure of the performance. The approach is also qualitatively compared against a Bayesian Nonparametric approach to identify the usefulness of the proposed method. Based on the analysis of results, there is a discussion about the usefulness and drawbacks of the method, followed by the scope for future research. / En utmaning att släppa autonoma fordon på allmänna vägar är säkerhetsverifiering av de utvecklade funktionerna. Säkerhetstestning av fordon är inte praktiskt genomförbart eftersom acceptanskriteriet kör minst 2,1 miljarder kilometer [1]. Ett alternativ till denna distansbaserade testning är det scenaribaserade tillväga-gångssättet, där intelligenta fordon utsätts för kända scenarier. Identifiering av sådana scenarier från kördata är avgörande för denna validering. Syftet med denna avhandling är att undersöka möjligheten till oövervakad identifiering av körscenarier från kördata. Uppgiften utförs i två huvuddelar. Den första är segmenteringen av tidsseriedrivdata genom att detektera ändringspunkter, följt av klustring av de tidigare erhållna segmenten. Tidsseriesegmentering närmar sig med en Deep Learningmetod, medan den andra uppgiften utförs med hjälp av tidsseriekluster. Arbetet innehåller också ett visuellt tillvägagångssätt för att validera tidsserierna, följt av ett kvantitativt mått på prestanda. Tillvägagångssättet jämförs också med en Bayesian icke-parametrisk metod för att identifiera användbarheten av den föreslagna metoden. Baserat på analysen av resultaten diskuteras metodens användbarhet och nackdelar, följt av möjligheten för framtida forskning.

Digitization of the energy industry, introduction of smart grids and increasing regulation of electricity consumption metering have resulted in vast amounts of electricity data. This data presents a unique opportunity to understand the electricity usage and to make it more efficient, reducing electricity consumption and carbon emissions. An important initial step in analyzing the data is to identify anomalies. In this thesis the problem of anomaly detection in electricity consumption series is addressed using four machine learning methods: density based spatial clustering for applications with noise (DBSCAN), local outlier factor (LOF), isolation forest (iForest) and one-class support vector machine (OC-SVM). In order to evaluate the methods synthetic anomalies were introduced to the electricity consumption series and the methods were then evaluated for the two anomaly types point anomaly and collective anomaly. In addition to electricity consumption data, features describing the prior consumption, outdoor temperature and date-time properties were included in the models. Results indicate that the addition of the temperature feature and the lag features generally impaired anomaly detection performance, while the inclusion of date-time features improved it. Of the four methods, OC-SVM was found to perform the best at detecting point anomalies, while LOF performed the best at detecting collective anomalies. In an attempt to improve the models' detection power the electricity consumption series were de-trended and de-seasonalized and the same experiments were carried out. The models did not perform better on the decomposed series than on the non-decomposed. / Digitaliseringen av elbranschen, införandet av smarta nät samt ökad reglering av elmätning har resulterat i stora mängder eldata. Denna data skapar en unik möjlighet att analysera och förstå fastigheters elförbrukning för att kunna effektivisera den. Ett viktigt inledande steg i analysen av denna data är att identifiera möjliga anomalier. I denna uppsats testas fyra olika maskininlärningsmetoder för detektering av anomalier i elförbrukningsserier: densitetsbaserad spatiell klustring för applikationer med brus (DBSCAN), lokal avvikelse-faktor (LOF), isoleringsskog (iForest) och en-klass stödvektormaskin (OC-SVM). För att kunna utvärdera metoderna infördes syntetiska anomalier i elförbrukningsserierna och de fyra metoderna utvärderades därefter för de två anomalityperna punktanomali och gruppanomali. Utöver elförbrukningsdatan inkluderades även variabler som beskriver tidigare elförbrukning, utomhustemperatur och tidsegenskaper i modellerna. Resultaten tyder på att tillägget av temperaturvariabeln och lag-variablerna i allmänhet försämrade modellernas prestanda, medan införandet av tidsvariablerna förbättrade den. Av de fyra metoderna visade sig OC-SVM vara bäst på att detektera punktanomalier medan LOF var bäst på att detektera gruppanomalier. I ett försök att förbättra modellernas detekteringsförmåga utfördes samma experiment efter att elförbrukningsserierna trend- och säsongsrensats. Modellerna presterade inte bättre på de rensade serierna än på de icke-rensade.

No description available.

<p dir="ltr">Semantic pairs, which consist of related entities or concepts, serve as the foundation for comprehending the meaning of language in both written and spoken forms. These pairs enable to grasp the nuances of relationships between words, phrases, or ideas, forming the basis for more advanced language tasks like entity recognition, sentiment analysis, machine translation, and question answering. They allow to infer causality, identify hierarchies, and connect ideas within a text, ultimately enhancing the depth and accuracy of automated language processing.</p><p dir="ltr">Nevertheless, the task of extracting semantic pairs from sentences poses a significant challenge, necessitating the relevance of syntactic dependency patterns (SDPs). Thankfully, semantic relationships exhibit adherence to distinct SDPs when connecting pairs of entities. Recognizing this fact underscores the critical importance of extracting these SDPs, particularly for specific semantic relationships like hyponym-hypernym, meronym-holonym, and cause-effect associations. The automated extraction of such SDPs carries substantial advantages for various downstream applications, including entity extraction, ontology development, and question answering. Unfortunately, this pivotal facet of pattern extraction has remained relatively overlooked by researchers in the domains of natural language processing (NLP) and information retrieval.</p><p dir="ltr">To address this gap, I introduce an attention-based supervised deep learning model, ASPER. ASPER is designed to extract SDPs that denote semantic relationships between entities within a given sentential context. I rigorously evaluate the performance of ASPER across three distinct semantic relations: hyponym-hypernym, cause-effect, and meronym-holonym, utilizing six datasets. My experimental findings demonstrate ASPER's ability to automatically identify an array of SDPs that mirror the presence of these semantic relationships within sentences, outperforming existing pattern extraction methods by a substantial margin.</p><p dir="ltr">Second, I want to use the SDPs to extract semantic pairs from sentences. I choose to extract cause-effect entities from medical literature. This task is instrumental in compiling various causality relationships, such as those between diseases and symptoms, medications and side effects, and genes and diseases. Existing solutions excel in sentences where cause and effect phrases are straightforward, such as named entities, single-word nouns, or short noun phrases. However, in the complex landscape of medical literature, cause and effect expressions often extend over several words, stumping existing methods, resulting in incomplete extractions that provide low-quality, non-informative, and at times, conflicting information. To overcome this challenge, I introduce an innovative unsupervised method for extracting cause and effect phrases, PatternCausality tailored explicitly for medical literature. PatternCausality employs a set of cause-effect dependency patterns as templates to identify the key terms within cause and effect phrases. It then utilizes a novel phrase extraction technique to produce comprehensive and meaningful cause and effect expressions from sentences. Experiments conducted on a dataset constructed from PubMed articles reveal that PatternCausality significantly outperforms existing methods, achieving a remarkable order of magnitude improvement in the F-score metric over the best-performing alternatives. I also develop various PatternCausality variants that utilize diverse phrase extraction methods, all of which surpass existing approaches. PatternCausality and its variants exhibit notable performance improvements in extracting cause and effect entities in a domain-neutral benchmark dataset, wherein cause and effect entities are confined to single-word nouns or noun phrases of one to two words.</p><p dir="ltr">Nevertheless, PatternCausality operates within an unsupervised framework and relies heavily on SDPs, motivating me to explore the development of a supervised approach. Although SDPs play a pivotal role in semantic relation extraction, pattern-based methodologies remain unsupervised, and the multitude of potential patterns within a language can be overwhelming. Furthermore, patterns do not consistently capture the broader context of a sentence, leading to the extraction of false-positive semantic pairs. As an illustration, consider the hyponym-hypernym pattern <i>the w of u</i> which can correctly extract semantic pairs for a sentence like <i>the village of Aasu</i> but fails to do so for the phrase <i>the moment of impact</i>. The root cause of this limitation lies in the pattern's inability to capture the nuanced meaning of words and phrases in a sentence and their contextual significance. These observations have spurred my exploration of a third model, DepBERT which constitutes a dependency-aware supervised transformer model. DepBERT's primary contribution lies in introducing the underlying dependency structure of sentences to a language model with the aim of enhancing token classification performance. To achieve this, I must first reframe the task of semantic pair extraction as a token classification problem. The DepBERT model can harness both the tree-like structure of dependency patterns and the masked language architecture of transformers, marking a significant milestone, as most large language models (LLMs) predominantly focus on semantics and word co-occurrence while neglecting the crucial role of dependency architecture.</p><p dir="ltr">In summary, my overarching contributions in this thesis are threefold. First, I validate the significance of the dependency architecture within various components of sentences and publish SDPs that incorporate these dependency relationships. Subsequently, I employ these SDPs in a practical medical domain to extract vital cause-effect pairs from sentences. Finally, my third contribution distinguishes this thesis by integrating dependency relations into a deep learning model, enhancing the understanding of language and the extraction of valuable semantic associations.</p>

[pt] Um data center normalmente possui grande quantidade de máquinas com diferentes configurações de hardware. Múltiplas aplicações são executadas e software e hardware são constantemente atualizados. Para evitar a interrupção de aplicações críticas, que podem causar grandes prejuízos financeiros, os administradores de sistemas devem identificar e corrigir as falhas o mais cedo possível. No entanto, a identificação de falhas em data centers de produção muitas vezes ocorre apenas quando as aplicações e serviços já estão indisponíveis. Entre as diferentes causas da detecção tardia de falhas estão o uso técnicas de monitoração baseadas apenas em thresholds. O aumento crescente na complexidade de aplicações que são constantemente atualizadas torna difícil a configuração de thresholds ótimos para cada métrica e servidor. Este trabalho propõe o uso de técnicas de detecção de anomalias no lugar de técnicas baseadas em thresholds. Uma anomalia é um comportamento do sistema que é incomum e significativamente diferente do comportamento normal anterior. Desenvolvemos um algoritmo para detecção de anomalias, chamado DASRS (Decreased Anomaly Score by Repeated Sequence) que analisa em tempo real as métricas coletadas de servidores de um data center de produção. O DASRS apresentou excelentes resultados de acurácia, compatível com os algoritmos do estado da arte, além de tempo de processamento e consumo de memória menores. Por esse motivo, o DASRS atende aos requisitos de processamento em tempo real de um grande volume de dados. / [en] A data center typically has a large number of machines with different hardware configurations. Multiple applications are executed and software and hardware are constantly updated. To avoid disruption of critical applications, which can cause significant financial loss, system administrators should identify and correct failures as early as possible. However, fault-detection in production data centers often occurs only when applications and services are already unavailable. Among the different causes of late fault-detection are the use of thresholds-only monitoring techniques. The increasing complexity of constantly updating applications makes it difficult to set optimal thresholds for each metric and server. This paper proposes the use of anomaly detection techniques in place of thresholds based techniques. An anomaly is a system behavior that is unusual and significantly different from the previous normal behavior. We have developed an anomaly detection algorithm called Decreased Anomaly Score by Repeated Sequence (DASRS) that analyzes real-time metrics collected from servers in a production data center. DASRS has showed excellent accuracy results, compatible with state-of-the-art algorithms, and reduced processing time and memory consumption. For this reason, DASRS meets the real-time processing requirements of a large volume of data.

<p>    </p> <p>The counting process is the fundamental of many real-world problems with event data. Poisson process, used as the background intensity of Hawkes process, is the most commonly used point process. The Hawkes process, a self-exciting point process fits to temporal event data, spatial-temporal event data, and event data with covariates. We study the Hawkes process that fits to heterogeneous drug overdose data via a novel semi-parametric approach. The counting process is also related to survival data based on the fact that they both study the occurrences of events over time. We fit a Cox model to temporal event data with a large corpus that is processed into high dimensional covariates. We study the significant features that influence the intensity of events. </p>

L’apprentissage supervisé de réseaux hiérarchiques à grande échelle connaît présentement un succès fulgurant. Malgré cette effervescence, l’apprentissage non-supervisé représente toujours, selon plusieurs chercheurs, un élément clé de l’Intelligence Artificielle, où les agents doivent apprendre à partir d’un nombre potentiellement limité de données. Cette thèse s’inscrit dans cette pensée et aborde divers sujets de recherche liés au problème d’estimation de densité par l’entremise des machines de Boltzmann (BM), modèles graphiques probabilistes au coeur de l’apprentissage profond. Nos contributions touchent les domaines de l’échantillonnage, l’estimation de fonctions de partition, l’optimisation ainsi que l’apprentissage de représentations invariantes. Cette thèse débute par l’exposition d’un nouvel algorithme d'échantillonnage adaptatif, qui ajuste (de fa ̧con automatique) la température des chaînes de Markov sous simulation, afin de maintenir une vitesse de convergence élevée tout au long de l’apprentissage. Lorsqu’utilisé dans le contexte de l’apprentissage par maximum de vraisemblance stochastique (SML), notre algorithme engendre une robustesse accrue face à la sélection du taux d’apprentissage, ainsi qu’une meilleure vitesse de convergence. Nos résultats sont présent ́es dans le domaine des BMs, mais la méthode est générale et applicable à l’apprentissage de tout modèle probabiliste exploitant l’échantillonnage par chaînes de Markov. Tandis que le gradient du maximum de vraisemblance peut-être approximé par échantillonnage, l’évaluation de la log-vraisemblance nécessite un estimé de la fonction de partition. Contrairement aux approches traditionnelles qui considèrent un modèle donné comme une boîte noire, nous proposons plutôt d’exploiter la dynamique de l’apprentissage en estimant les changements successifs de log-partition encourus à chaque mise à jour des paramètres. Le problème d’estimation est reformulé comme un problème d’inférence similaire au filtre de Kalman, mais sur un graphe bi-dimensionnel, où les dimensions correspondent aux axes du temps et au paramètre de température. Sur le thème de l’optimisation, nous présentons également un algorithme permettant d’appliquer, de manière efficace, le gradient naturel à des machines de Boltzmann comportant des milliers d’unités. Jusqu’à présent, son adoption était limitée par son haut coût computationel ainsi que sa demande en mémoire. Notre algorithme, Metric-Free Natural Gradient (MFNG), permet d’éviter le calcul explicite de la matrice d’information de Fisher (et son inverse) en exploitant un solveur linéaire combiné à un produit matrice-vecteur efficace. L’algorithme est prometteur: en terme du nombre d’évaluations de fonctions, MFNG converge plus rapidement que SML. Son implémentation demeure malheureusement inefficace en temps de calcul. Ces travaux explorent également les mécanismes sous-jacents à l’apprentissage de représentations invariantes. À cette fin, nous utilisons la famille de machines de Boltzmann restreintes “spike & slab” (ssRBM), que nous modifions afin de pouvoir modéliser des distributions binaires et parcimonieuses. Les variables latentes binaires de la ssRBM peuvent être rendues invariantes à un sous-espace vectoriel, en associant à chacune d’elles, un vecteur de variables latentes continues (dénommées “slabs”). Ceci se traduit par une invariance accrue au niveau de la représentation et un meilleur taux de classification lorsque peu de données étiquetées sont disponibles. Nous terminons cette thèse sur un sujet ambitieux: l’apprentissage de représentations pouvant séparer les facteurs de variations présents dans le signal d’entrée. Nous proposons une solution à base de ssRBM bilinéaire (avec deux groupes de facteurs latents) et formulons le problème comme l’un de “pooling” dans des sous-espaces vectoriels complémentaires. / Despite the current widescale success of deep learning in training large scale hierarchical models through supervised learning, unsupervised learning promises to play a crucial role towards solving general Artificial Intelligence, where agents are expected to learn with little to no supervision. The work presented in this thesis tackles the problem of unsupervised feature learning and density estimation, using a model family at the heart of the deep learning phenomenon: the Boltzmann Machine (BM). We present contributions in the areas of sampling, partition function estimation, optimization and the more general topic of invariant feature learning. With regards to sampling, we present a novel adaptive parallel tempering method which dynamically adjusts the temperatures under simulation to maintain good mixing in the presence of complex multi-modal distributions. When used in the context of stochastic maximum likelihood (SML) training, the improved ergodicity of our sampler translates to increased robustness to learning rates and faster per epoch convergence. Though our application is limited to BM, our method is general and is applicable to sampling from arbitrary probabilistic models using Markov Chain Monte Carlo (MCMC) techniques. While SML gradients can be estimated via sampling, computing data likelihoods requires an estimate of the partition function. Contrary to previous approaches which consider the model as a black box, we provide an efficient algorithm which instead tracks the change in the log partition function incurred by successive parameter updates. Our algorithm frames this estimation problem as one of filtering performed over a 2D lattice, with one dimension representing time and the other temperature. On the topic of optimization, our thesis presents a novel algorithm for applying the natural gradient to large scale Boltzmann Machines. Up until now, its application had been constrained by the computational and memory requirements of computing the Fisher Information Matrix (FIM), which is square in the number of parameters. The Metric-Free Natural Gradient algorithm (MFNG) avoids computing the FIM altogether by combining a linear solver with an efficient matrix-vector operation. The method shows promise in that the resulting updates yield faster per-epoch convergence, despite being slower in terms of wall clock time. Finally, we explore how invariant features can be learnt through modifications to the BM energy function. We study the problem in the context of the spike & slab Restricted Boltzmann Machine (ssRBM), which we extend to handle both binary and sparse input distributions. By associating each spike with several slab variables, latent variables can be made invariant to a rich, high dimensional subspace resulting in increased invariance in the learnt representation. When using the expected model posterior as input to a classifier, increased invariance translates to improved classification accuracy in the low-label data regime. We conclude by showing a connection between invariance and the more powerful concept of disentangling factors of variation. While invariance can be achieved by pooling over subspaces, disentangling can be achieved by learning multiple complementary views of the same subspace. In particular, we show how this can be achieved using third-order BMs featuring multiplicative interactions between pairs of random variables.