Unnormalised parametric models are an important class of probabilistic models which are difficult to estimate. The models are important since they occur in many different areas of application, e.g. in modelling of natural images, natural language and associative memory. However, standard maximum likelihood estimation is not applicable to unnormalised models, so alternative methods are required. Noise-contrastive estimation (NCE) has been proposed as an effective estimation method for unnormalised models. The basic idea is to transform the unsupervised estimation problem into a supervised classification problem. The parameters of the unnormalised model are learned by training the model to differentiate the given data samples from generated noise samples. However, the choice of the noise distribution has been left open to the user, and as the performance of the estimation may be sensitive to this choice, it is desirable for it to be automated. In this thesis, the ambiguity in the choice of the noise distribution is addressed by presenting the previously unpublished conditional noise-contrastive estimation (CNCE) method. Like NCE, CNCE estimates unnormalised models by classifying data and noise samples. However, the choice of noise distribution is partly automated via the use of a conditional noise distribution that is dependent on the data. In addition to introducing the core theory for CNCE, the method is empirically validated on data and models where the ground truth is known. Furthermore, CNCE is applied to natural image data to show its applicability in a realistic application. / Icke-normaliserade parametriska modeller utgör en viktig klass av svåruppskattade statistiska modeller. Dessa modeller är viktiga eftersom de uppträder inom många olika tillämpningsområden, t.ex. vid modellering av bilder, tal och skrift och associativt minne. Dessa modeller är svåruppskattade eftersom den vanliga maximum likelihood-metoden inte är tillämpbar på icke-normaliserade modeller. Noise-contrastive estimation (NCE) har föreslagits som en effektiv metod för uppskattning av icke-normaliserade modeller. Grundidén är att transformera det icke-handledda uppskattningsproblemet till ett handlett klassificeringsproblem. Den icke-normaliserade modellens parametrar blir inlärda genom att träna modellen på att skilja det givna dataprovet från ett genererat brusprov. Dock har valet av brusdistribution lämnats öppet för användaren. Eftersom uppskattningens prestanda är känslig gentemot det här valet är det önskvärt att få det automatiserat. I det här examensarbetet behandlas valet av brusdistribution genom att presentera den tidigare opublicerade metoden conditional noise-contrastive estimation (CNCE). Liksom NCE uppskattar CNCE icke-normaliserade modeller via klassificering av data- och brusprov. I det här fallet är emellertid brusdistributionen delvis automatiserad genom att använda en betingad brusdistribution som är beroende på dataprovet. Förutom att introducera kärnteorin för CNCE valideras även metoden med hjälp av data och modeller vars genererande parametrar är kända. Vidare appliceras CNCE på bilddata för att demonstrera dess tillämpbarhet.

Durant ces dernières années, les architectures de réseaux de neurones (RN) ont été appliquées avec succès à de nombreuses applications en Traitement Automatique de Langues (TAL), comme par exemple en Reconnaissance Automatique de la Parole (RAP) ainsi qu'en Traduction Automatique (TA).Pour la tâche de modélisation statique de la langue, ces modèles considèrent les unités linguistiques (c'est-à-dire des mots et des segments) à travers leurs projections dans un espace continu (multi-dimensionnel), et la distribution de probabilité à estimer est une fonction de ces projections.Ainsi connus sous le nom de "modèles continus" (MC), la particularité de ces derniers se trouve dans l'exploitation de la représentation continue qui peut être considérée comme une solution au problème de données creuses rencontré lors de l'utilisation des modèles discrets conventionnels.Dans le cadre de la TA, ces techniques ont été appliquées dans les modèles de langue neuronaux (MLN) utilisés dans les systèmes de TA, et dans les modèles continus de traduction (MCT).L'utilisation de ces modèles se sont traduit par d'importantes et significatives améliorations des performances des systèmes de TA. Ils sont néanmoins très coûteux lors des phrases d'apprentissage et d'inférence, notamment pour les systèmes ayant un grand vocabulaire.Afin de surmonter ce problème, l'architecture SOUL (pour "Structured Output Layer" en anglais) et l'algorithme NCE (pour "Noise Contrastive Estimation", ou l'estimation contrastive bruitée) ont été proposés: le premier modifie la structure standard de la couche de sortie, alors que le second cherche à approximer l'estimation du maximum de vraisemblance (MV) par une méthode d’échantillonnage.Toutes ces approches partagent le même critère d'estimation qui est la log-vraisemblance; pourtant son utilisation mène à une incohérence entre la fonction objectif définie pour l'estimation des modèles, et la manière dont ces modèles seront utilisés dans les systèmes de TA.Cette dissertation vise à concevoir de nouvelles procédures d'entraînement des MC, afin de surmonter ces problèmes.Les contributions principales se trouvent dans l'investigation et l'évaluation des méthodes d'entraînement efficaces pour MC qui visent à: (i) réduire le temps total de l'entraînement, et (ii) améliorer l'efficacité de ces modèles lors de leur utilisation dans les systèmes de TA.D'un côté, le coût d'entraînement et d'inférence peut être réduit (en utilisant l'architecture SOUL ou l'algorithme NCE), ou la convergence peut être accélérée.La dissertation présente une analyse empirique de ces approches pour des tâches de traduction automatique à grande échelle.D'un autre côté, nous proposons un cadre d'apprentissage discriminant qui optimise la performance du système entier ayant incorporé un modèle continu.Les résultats expérimentaux montrent que ce cadre d'entraînement est efficace pour l'apprentissage ainsi que pour l'adaptation des MC au sein des systèmes de TA, ce qui ouvre de nouvelles perspectives prometteuses. / Over the past few years, neural network (NN) architectures have been successfully applied to many Natural Language Processing (NLP) applications, such as Automatic Speech Recognition (ASR) and Statistical Machine Translation (SMT).For the language modeling task, these models consider linguistic units (i.e words and phrases) through their projections into a continuous (multi-dimensional) space, and the estimated distribution is a function of these projections. Also qualified continuous-space models (CSMs), their peculiarity hence lies in this exploitation of a continuous representation that can be seen as an attempt to address the sparsity issue of the conventional discrete models. In the context of SMT, these echniques have been applied on neural network-based language models (NNLMs) included in SMT systems, and oncontinuous-space translation models (CSTMs). These models have led to significant and consistent gains in the SMT performance, but are also considered as very expensive in training and inference, especially for systems involving large vocabularies. To overcome this issue, Structured Output Layer (SOUL) and Noise Contrastive Estimation (NCE) have been proposed; the former modifies the standard structure on vocabulary words, while the latter approximates the maximum-likelihood estimation (MLE) by a sampling method. All these approaches share the same estimation criterion which is the MLE ; however using this procedure results in an inconsistency between theobjective function defined for parameter stimation and the way models are used in the SMT application. The work presented in this dissertation aims to design new performance-oriented and global training procedures for CSMs to overcome these issues. The main contributions lie in the investigation and evaluation of efficient training methods for (large-vocabulary) CSMs which aim~:(a) to reduce the total training cost, and (b) to improve the efficiency of these models when used within the SMT application. On the one hand, the training and inference cost can be reduced (using the SOUL structure or the NCE algorithm), or by reducing the number of iterations via a faster convergence. This thesis provides an empirical analysis of these solutions on different large-scale SMT tasks. On the other hand, we propose a discriminative training framework which optimizes the performance of the whole system containing the CSM as a component model. The experimental results show that this framework is efficient to both train and adapt CSM within SMT systems, opening promising research perspectives.

Automotive technologies and fully autonomous driving have seen a tremendous growth in recent times and have benefitted from extensive deep learning research. State-of-the-art deep learning methods are largely supervised and require labelled data for training. However, the annotation process for image data is time-consuming and costly in terms of human efforts. It is of interest to find informative samples for labelling by Content Based Image Retrieval (CBIR). Generally, a CBIR method takes a query image as input and returns a set of images that are semantically similar to the query image. The image retrieval is achieved by transforming images to feature representations in a latent space, where it is possible to reason about image similarity in terms of image content. In this thesis, a self-supervised method is developed to learn feature representations of road scenes images. The self-supervised method learns feature representations for images by adapting intermediate convolutional features from an existing deep Convolutional Neural Network (CNN). A contrastive approach based on Noise Contrastive Estimation (NCE) is used to train the feature learning model. For complex images like road scenes where mutiple image aspects can occur simultaneously, it is important to embed all the salient image aspects in the feature representation. To achieve this, the output feature representation is obtained as an ensemble of feature embeddings which are learned by focusing on different image aspects. An attention mechanism is incorporated to encourage each ensemble member to focus on different image aspects. For comparison, a self-supervised model without attention is considered and a simple dimensionality reduction approach using SVD is treated as the baseline. The methods are evaluated on nine different evaluation datasets using CBIR performance metrics. The datasets correspond to different image aspects and concern the images at different spatial levels - global, semi-global and local. The feature representations learned by self-supervised methods are shown to perform better than the SVD approach. Taking into account that no labelled data is required for training, learning representations for road scenes images using self-supervised methods appear to be a promising direction. Usage of multiple query images to emphasize a query intention is investigated and a clear improvement in CBIR performance is observed. It is inconclusive whether the addition of an attentive mechanism impacts CBIR performance. The attention method shows some positive signs based on qualitative analysis and also performs better than other methods for one of the evaluation datasets containing a local aspect. This method for learning feature representations is promising but requires further research involving more diverse and complex image aspects.