Nous proposons dans cette thèse de développer une méthode de restauration aveugle d'images flouées et bruitées où aucune connaissance a priori n'est exigée. Ce manuscrit est composé de trois chapitres : le 1er chapitre est consacré aux travaux de l'état de l'art. Les approches d'optimisation pour la résolution du problème de restauration y sont d'abord discutées. Ensuite les principales méthodes de restauration, dites semi-aveugles car nécessitant un minimum de connaissance a priori sont analysées. Parmi ces méthodes, cinq sont retenues pour évaluation. Le 2ème chapitre est dédié à la comparaison des performances des méthodes retenues dans le chapitre précédent. Les principaux critères objectifs d'évaluation de la qualité des images restaurées sont présentés. Parmi ces critères, la norme L1 de l'erreur d'estimation est sélectionnée. L'étude comparative menée sur une banque d'images monochromes, dégradées artificiellement par deux fonctions floues de supports différents et trois niveaux de bruit a permis de mettre en évidence les deux méthodes les plus pertinentes. La première repose sur une approche alternée mono-échelle où la PSF et l'image sont estimées dans une seule étape. La seconde utilise une approche hybride multi-échelle qui consiste tout d'abord à estimer de manière alternée la PSF et une image latente, puis dans une étape suivante séquentielle, à restaurer l'image. Dans l'étude comparative conduite, l'avantage revient à cette dernière. Les performances de ces méthodes serviront de référence pour comparer ensuite la méthode développée. Le 3ème chapitre porte sur la méthode développée. Nous avons cherché à rendre aveugle l'approche hybride retenue dans le chapitre précédent tout en améliorant la qualité d'estimation de la PSF et de l'image restaurée. Les contributions ont porté sur plusieurs points. Une première série d'améliorations concerne la redéfinition des échelles, celle de l'initialisation de l'image latente à chaque niveau d'échelle, l'évolution des paramètres pour la sélection des contours pertinents servant de support à l'estimation de la PSF et enfin, la définition d'un critère d'arrêt aveugle. Une seconde série de contributions a porté sur l'estimation aveugle des deux paramètres de régularisation impliqués pour éviter d'avoir à les fixer empiriquement. Chaque paramètre est associé à une fonction coût distincte l'une pour l'estimation de la PSF et la seconde pour l'estimation d'une image latente. Dans l'étape séquentielle qui suit, nous avons cherché à affiner le support de la PSF estimée dans l'étape alternée, avant de l'exploiter dans le processus de restauration de l'image. A ce niveau, la seule connaissance a priori nécessaire est une borne supérieure du support de la PSF. Les différentes évaluations conduites sur des images monochromes et hyperspectrales dégradées artificiellement par plusieurs flous de type mouvement, de supports différents, montrent une nette amélioration de la qualité de restauration obtenue par l'approche développée par rapport aux deux meilleures approches de l'état de l'art retenues. / We propose in this thesis manuscript to develop a blind restoration method of single component blurred and noisy images where no prior knowledge is required. This manuscript is composed of three chapters: the first chapter focuses on state-of-art works. The optimization approaches for resolving the restoration problem are discussed first. Then, the main methods of restoration, so-called semi-blind ones because requiring a minimum of a priori knowledge are analysed. Five of these methods are selected for evaluation. The second chapter is devoted to comparing the performance of the methods selected in the previous chapter. The main objective criteria for evaluating the quality of the restored images are presented. Of these criteria, the l1 norm for the estimation error is selected. The comparative study conducted on a database of monochromatic images, artificially degraded by two blurred functions with different support size and three levels of noise, revealed the most two relevant methods. The first one is based on a single-scale alternating approach where both the PSF and the image are estimated alternatively. The second one uses a multi-scale hybrid approach, which consists first of alternatingly estimating the PSF and a latent image, then in a sequential next step, restoring the image. In the comparative study performed, the benefit goes to the latter. The performance of both these methods will be used as references to then compare the newly designed method. The third chapter deals with the developed method. We have sought to make the hybrid approach retained in the previous chapter as blind as possible while improving the quality of estimation of both the PSF and the restored image. The contributions covers a number of points. A first series concerns the redefinition of the scales that of the initialization of the latent image at each scale level, the evolution of the parameters for the selection of the relevant contours supporting the estimation of the PSF and finally the definition of a blind stop criterion. A second series of contributions concentrates on the blind estimation of the two regularization parameters involved in order to avoid having to fix them empirically. Each parameter is associated with a separate cost function either for the PSF estimation or for the estimation of a latent image. In the sequential step that follows, we refine the estimation of the support of the PSF estimated in the previous alternated step, before exploiting it in the process of restoring the image. At this level, the only a priori knowledge necessary is a higher bound of the support of the PSF. The different evaluations performed on monochromatic and hyperspectral images artificially degraded by several motion-type blurs with different support sizes, show a clear improvement in the quality of restoration obtained by the newly designed method in comparison to the best two state-of-the-art methods retained.

Cette thèse concerne le problème de désocclusion d'images, au moyen des équations aux dérivées partielles. Dans la première partie de la thèse, la désocclusion est modélisée par un problème de Cauchy qui consiste à déterminer une solution d'une équation aux dérivées partielles avec des données aux bords accessibles seulement sur une partie du bord de la partie à recouvrir. Ensuite, on a utilisé des algorithmes de minimisation issus de la théorie des jeux, pour résoudre ce problème de Cauchy. La deuxième partie de la thèse est consacrée au choix des paramètres de régularisation pour des EDP d'ordre deux et d'ordre quatre. L'approche développée consiste à construire une famille de problèmes d'optimisation bien posés où les paramètres sont choisis comme étant une fonction variable en espace. Ceci permet de prendre en compte les différents détails, à différents échelles dans l'image. L'apport de la méthode est de résoudre de façon satisfaisante et objective, le choix du paramètre de régularisation en se basant sur des indicateurs d'erreur et donc le caractère à posteriori de la méthode (i.e. indépendant de la solution exacte, en générale inconnue). En outre, elle fait appel à des techniques classiques d'adaptation de maillage, qui rendent peu coûteuses les calculs numériques. En plus, un des aspects attractif de cette méthode, en traitement d'images est la récupération et la détection de contours et de structures fines. / Image inpainting refers to the process of restoring a damaged image with missing information. Different mathematical approaches were suggested to deal with this problem. In particular, partial differential diffusion equations are extensively used. The underlying idea of PDE-based approaches is to fill-in damaged regions with available information from their surroundings. The first purpose of this Thesis is to treat the case where this information is not available in a part of the boundary of the damaged region. We formulate the inpainting problem as a nonlinear boundary inverse problem for incomplete images. Then, we give a Nash-game formulation of this Cauchy problem and we present different numerical which show the efficiency of the proposed approach as an inpainting method.Typically, inpainting is an ill-posed inverse problem for it most of PDEs approaches are obtained from minimization of regularized energies, in the context of Tikhonov regularization. The second part of the thesis is devoted to the choice of regularization parameters in second-and fourth-order energy-based models with the aim of obtaining as far as possible fine features of the initial image, e.g., (corners, edges, … ) in the inpainted region. We introduce a family of regularized functionals with regularization parameters to be selected locally, adaptively and in a posteriori way allowing to change locally the initial model. We also draw connections between the proposed method and the Mumford-Shah functional. An important feature of the proposed method is that the investigated PDEs are easy to discretize and the overall adaptive approach is easy to implement numerically.

Grâce aux progrès des techniques de séquençage, les bases de données génomiques ont connu une croissance exponentielle depuis la fin des années 1990. Un grand nombre d'outils statistiques ont été développés à l'interface entre bioinformatique, apprentissage automatique et physique statistique, dans le but d'extraire de l'information de ce déluge de données. Plusieurs approches de physique statistique ont été récemment introduites dans le contexte précis de la modélisation de séquences de protéines, dont l'analyse en couplages directs. Cette méthode d'inférence statistique globale fondée sur le principe d'entropie maximale, s'est récemment montrée d'une efficacité redoutable pour prédire la structure tridimensionnelle de protéines, à partir de considérations purement statistiques.Dans cette thèse, nous présentons les méthodes d'inférence en question, et encouragés par leur succès, explorons d'autres domaines complexes dans lesquels elles pourraient être appliquées, comme la détection d'homologies. Contrairement à la prédiction des contacts entre résidus qui se limite à une information topologique sur le réseau d'interactions, ces nouveaux champs d'application exigent des considérations énergétiques globales et donc un modèle plus quantitatif et détaillé. À travers une étude approfondie sur des donnéesartificielles et biologiques, nous proposons une meilleure interpretation des paramètres centraux de ces méthodes d'inférence, jusqu'ici mal compris, notamment dans le cas d'un échantillonnage limité. Enfin, nous présentons une nouvelle procédure plus précise d'inférence de modèles génératifs, qui mène à des avancées importantes pour des données réelles en quantité limitée. / Over the last decades, genomic databases have grown exponentially in size thanks to the constant progress of modern DNA sequencing. A large variety of statistical tools have been developed, at the interface between bioinformatics, machine learning, and statistical physics, to extract information from these ever increasing datasets. In the specific context of protein sequence data, several approaches have been recently introduced by statistical physicists, such as direct-coupling analysis, a global statistical inference method based on the maximum-entropy principle, that has proven to be extremely effective in predicting the three-dimensional structure of proteins from purely statistical considerations.In this dissertation, we review the relevant inference methods and, encouraged by their success, discuss their extension to other challenging fields, such as sequence folding prediction and homology detection. Contrary to residue-residue contact prediction, which relies on an intrinsically topological information about the network of interactions, these fields require global energetic considerations and therefore a more quantitative and detailed model. Through an extensive study on both artificial and biological data, we provide a better interpretation of the central inferred parameters, up to now poorly understood, especially in the limited sampling regime. Finally, we present a new and more precise procedure for the inference of generative models, which leads to further improvements on real, finitely sampled data.

Cette thèse est consacrée à la recherche des zéros d'un opérateur maximal monotone structuré, à l'aide de systèmes dynamiques dissipatifs continus et discrets. Les solutions sont obtenues comme limites des trajectoires lorsque le temps t tend vers l'infini. On s'intéressera principalement aux dynamiques obtenues par régularisation de type Levenberg-Marquardt de la méthode de Newton. On décrira aussi les approches basées sur des dynamiques voisines.Dans un cadre Hilbertien, on s'intéresse à la recherche des zéros de l'opérateur maximal monotone structuré M = A + B, où A est un opérateur maximal monotone général et B est un opérateur monotone Lipschitzien. Nous introduisons des dynamiques continues et discrètes de type Newton régularisé faisant intervenir d'une façon séparée les résolvantes de l'opérateur A (implicites), et des évaluations de B (explicites). A l'aide de la représentation de Minty de l'opérateur A comme une variété Lipschitzienne, nous reformulons ces dynamiques sous une forme relevant du théorème de Cauchy-Lipschitz. Nous nous intéressons au cas particulier où A est le sous différentiel d'une fonction convexe, semi-continue inférieurement, et propre, et B est le gradient d'une fonction convexe, différentiable. Nous étudions le comportement asymptotique des trajectoires. Lorsque le terme de régularisation ne tend pas trop vite vers zéro, et en s'appuyant sur une analyse asymptotique de type Lyapunov, nous montrons la convergence des trajectoires. Par ailleurs, nous montrons la dépendance Lipschitzienne des trajectoires par rapport au terme de régularisation.Puis nous élargissons notre étude en considérant différentes classes de systèmes dynamiques visant à résoudre les inclusions monotones gouvernées par un opérateur maximal monotone structuré M = $partialPhi$+ B, où $partialPhi$ désigne le sous différentiel d'une fonction convexe, semicontinue inférieurement, et propre, et B est un opérateur monotone cocoercif. En s'appuyant sur une analyse asymptotique de type Lyapunov, nous étudions le comportement asymptotique des trajectoires de ces systèmes. La discrétisation temporelle de ces dynamiques fournit desalgorithmes forward-backward (certains nouveaux ).Finalement, nous nous intéressons à l'étude du comportement asymptotique des trajectoires de systèmes dynamiques de type Newton régularisé, dans lesquels on introduit un terme supplémentaire de viscosité évanescente de type Tikhonov. On obtient ainsi la sélection asymptotique d'une solution de norme minimale. / This thesis is devoted to finding zeroes of structured maximal monotone operators, by using discrete and continuous dissipative dynamical systems. The solutions are obtained as the limits of trajectories when the time t tends towards infinity.We pay special attention to the dynamics that are obtained by Levenberg-Marquardt regularization of Newton's method. We also revisit the approaches based on some related dynamical systems.In a Hilbert framework, we are interested in finding zeroes of a structured maximal monotone operator M = A + B, where A is a general maximal monotone operator, and B is monotone and locally Lipschitz continuous. We introduce discrete and continuous dynamical systems which are linked to Newton's method. They involve separately B and the resolvents of A, and are designed to splitting methods. Based on the Minty representation of A as a Lipschitz manifold, we show that these dynamics can be formulated as differential systems, which are relevant to the Cauchy-Lipschitz theorem. We focus on the particular case where A is the subdifferential of a convex lower semicontinuous proper function, and B is the gradient of a convex, continuously differentiable function. We study the asymptotic behavior of trajectories. When the regularization parameter does not tend to zero too rapidly, and by using Lyapunov asymptotic analysis, we show the convergence of trajectories. Besides, we show the Lipschitz continuous dependence of the solution with respect to the regularization term.Then we extend our study by considering various classes of dynamical systems which aim at solving inclusions governed by structured monotone operators M = $partialPhi$+ B, where $partialPhi$ is the subdifferential of a convex lower semicontinuous function, and B is a monotone cocoercive operator. By a Lyapunov analysis, we show the convergence properties of the orbits of these systems. The time discretization of these dynamics gives various forward-backward splittingmethods (some new).Finally, we focus on the study of the asymptotic behavior of trajectories of the regularized Newton dynamics, in which we introduce an additional vanishing Tikhonov-like viscosity term.We thus obtain the asymptotic selection of the solution of minimal norm.

Artificial neural networks have been widely used for machine learning and optimization. A neuro ensemble is a collection of neural networks that works cooperatively on a problem. In the literature, it has been shown that by combining several neural networks, the generalization of the overall system could be enhanced over the separate generalization ability of the individuals. Evolutionary computation can be used to search for a suitable architecture and weights for neural networks. When evolutionary computation is used to evolve a neuro ensemble, it is usually known as evolutionary neuro ensemble. In most real-world problems, we either know little about these problems or the problems are too complex to have a clear vision on how to decompose them by hand. Thus, it is usually desirable to have a method to automatically decompose a complex problem into a set of overlapping or non-overlapping sub-problems and assign one or more specialists (i.e. experts, learning machines) to each of these sub-problems. An important feature of neuro ensemble is automatic problem decomposition. Some neuro ensemble methods are able to generate networks, where each individual network is specialized on a unique sub-task such as mapping a subspace of the feature space. In real world problems, this is usually an important feature for a number of reasons including: (1) it provides an understanding of the decomposition nature of a problem; (2) if a problem changes, one can replace the network associated with the sub-space where the change occurs without affecting the overall ensemble; (3) if one network fails, the rest of the ensemble can still function in their sub-spaces; (4) if one learn the structure of one problem, it can potentially be transferred to other similar problems. In this thesis, I focus on classification problems and present a systematic study of a novel evolutionary neuro ensemble approach which I call cooperative coevolutionary mixture of experts (CCME). Cooperative coevolution (CC) is a branch of evolutionary computation where individuals in different populations cooperate to solve a problem and their fitness function is calculated based on their reciprocal interaction. The mixture of expert model (ME) is a neuro ensemble approach which can generate networks that are specialized on different sub-spaces in the feature space. By combining CC and ME, I have a powerful framework whereby it is able to automatically form the experts and train each of them. I show that the CCME method produces competitive results in terms of generalization ability without increasing the computational cost when compared to traditional training approaches. I also propose two different mechanisms for visualizing the resultant decomposition in high-dimensional feature spaces. The first mechanism is a simple one where data are grouped based on the specialization of each expert and a color-map of the data records is visualized. The second mechanism relies on principal component analysis to project the feature space onto lower dimensions, whereby decision boundaries generated by each expert are visualized through convex approximations. I also investigate the regularization effect of learning by forgetting on the proposed CCME. I show that learning by forgetting helps CCME to generate neuro ensembles of low structural complexity while maintaining their generalization abilities. Overall, the thesis presents an evolutionary neuro ensemble method whereby (1) the generated ensemble generalizes well; (2) it is able to automatically decompose the classification problem; and (3) it generates networks with small architectures.

In recent years, probabilistic models have become fundamental techniques in machine learning. They are successfully applied in various engineering problems, such as robotics, biometrics, brain-computer interfaces or artificial vision, and will gain in importance in the near future. This work deals with the difficult, but common situation where the data is, either very noisy, or scarce compared to the complexity of the process to model. We focus on latent variable models, which can be formalized as probabilistic graphical models and learned by the expectation-maximization algorithm or its variants (e.g., variational Bayes).<br> After having carefully studied a non-exhaustive list of multivariate kernel density estimators, we established that in most applications locally adaptive estimators should be preferred. Unfortunately, these methods are usually sensitive to outliers and have often too many parameters to set. Therefore, we focus on finite mixture models, which do not suffer from these drawbacks provided some structural modifications.<br> Two questions are central in this dissertation: (i) how to make mixture models robust to noise, i.e. deal efficiently with outliers, and (ii) how to exploit side-channel information, i.e. additional information intrinsic to the data. In order to tackle the first question, we extent the training algorithms of the popular Gaussian mixture models to the Student-t mixture models. the Student-t distribution can be viewed as a heavy-tailed alternative to the Gaussian distribution, the robustness being tuned by an extra parameter, the degrees of freedom. Furthermore, we introduce a new variational Bayesian algorithm for learning Bayesian Student-t mixture models. This algorithm leads to very robust density estimators and clustering. To address the second question, we introduce manifold constrained mixture models. This new technique exploits the information that the data is living on a manifold of lower dimension than the dimension of the feature space. Taking the implicit geometrical data arrangement into account results in better generalization on unseen data.<br> Finally, we show that the latent variable framework used for learning mixture models can be extended to construct probabilistic regularization networks, such as the Relevance Vector Machines. Subsequently, we make use of these methods in the context of an optic nerve visual prosthesis to restore partial vision to blind people of whom the optic nerve is still functional. Although visual sensations can be induced electrically in the blind's visual field, the coding scheme of the visual information along the visual pathways is poorly known. Therefore, we use probabilistic models to link the stimulation parameters to the features of the visual perceptions. Both black-box and grey-box models are considered. The grey-box models take advantage of the known neurophysiological information and are more instructive to medical doctors and psychologists.<br>

In dieser Arbeit werden statistische Lernprobleme betrachtet. Lernmaschinen extrahieren Informationen aus einer gegebenen Menge von Trainingsmustern, so daß sie in der Lage sind, Eigenschaften von bisher ungesehenen Mustern - z.B. eine Klassenzugehörigkeit - vorherzusagen. Wir betrachten den Fall, bei dem die resultierende Klassifikations- oder Regressionsregel aus einfachen Regeln - den Basishypothesen - zusammengesetzt ist. Die sogenannten Boosting Algorithmen erzeugen iterativ eine gewichtete Summe von Basishypothesen, die gut auf ungesehenen Mustern vorhersagen. <br /> Die Arbeit behandelt folgende Sachverhalte: <br /> <br /> o Die zur Analyse von Boosting-Methoden geeignete Statistische Lerntheorie. Wir studieren lerntheoretische Garantien zur Abschätzung der Vorhersagequalität auf ungesehenen Mustern. Kürzlich haben sich sogenannte Klassifikationstechniken mit großem Margin als ein praktisches Ergebnis dieser Theorie herausgestellt - insbesondere Boosting und Support-Vektor-Maschinen. Ein großer Margin impliziert eine hohe Vorhersagequalität der Entscheidungsregel. Deshalb wird analysiert, wie groß der Margin bei Boosting ist und ein verbesserter Algorithmus vorgeschlagen, der effizient Regeln mit maximalem Margin erzeugt.<br /> <br /> o Was ist der Zusammenhang von Boosting und Techniken der konvexen Optimierung? <br /> Um die Eigenschaften der entstehenden Klassifikations- oder Regressionsregeln zu analysieren, ist es sehr wichtig zu verstehen, ob und unter welchen Bedingungen iterative Algorithmen wie Boosting konvergieren. Wir zeigen, daß solche Algorithmen benutzt werden koennen, um sehr große Optimierungsprobleme mit Nebenbedingungen zu lösen, deren Lösung sich gut charakterisieren laesst. Dazu werden Verbindungen zum Wissenschaftsgebiet der konvexen Optimierung aufgezeigt und ausgenutzt, um Konvergenzgarantien für eine große Familie von Boosting-ähnlichen Algorithmen zu geben.<br /> <br /> o Kann man Boosting robust gegenüber Meßfehlern und Ausreissern in den Daten machen? <br /> Ein Problem bisheriger Boosting-Methoden ist die relativ hohe Sensitivität gegenüber Messungenauigkeiten und Meßfehlern in der Trainingsdatenmenge. Um dieses Problem zu beheben, wird die sogenannte 'Soft-Margin' Idee, die beim Support-Vector Lernen schon benutzt wird, auf Boosting übertragen. Das führt zu theoretisch gut motivierten, regularisierten Algorithmen, die ein hohes Maß an Robustheit aufweisen.<br /> <br /> o Wie kann man die Anwendbarkeit von Boosting auf Regressionsprobleme erweitern? <br /> Boosting-Methoden wurden ursprünglich für Klassifikationsprobleme entwickelt. Um die Anwendbarkeit auf Regressionsprobleme zu erweitern, werden die vorherigen Konvergenzresultate benutzt und neue Boosting-ähnliche Algorithmen zur Regression entwickelt. Wir zeigen, daß diese Algorithmen gute theoretische und praktische Eigenschaften haben.<br /> <br /> o Ist Boosting praktisch anwendbar? <br /> Die dargestellten theoretischen Ergebnisse werden begleitet von Simulationsergebnissen, entweder, um bestimmte Eigenschaften von Algorithmen zu illustrieren, oder um zu zeigen, daß sie in der Praxis tatsächlich gut funktionieren und direkt einsetzbar sind. Die praktische Relevanz der entwickelten Methoden wird in der Analyse chaotischer Zeitreihen und durch industrielle Anwendungen wie ein Stromverbrauch-Überwachungssystem und bei der Entwicklung neuer Medikamente illustriert. / In this work we consider statistical learning problems. A learning machine aims to extract information from a set of training examples such that it is able to predict the associated label on unseen examples. We consider the case where the resulting classification or regression rule is a combination of simple rules - also called base hypotheses. The so-called boosting algorithms iteratively find a weighted linear combination of base hypotheses that predict well on unseen data. We address the following issues:<br /> <br /> o The statistical learning theory framework for analyzing boosting methods.<br /> We study learning theoretic guarantees on the prediction performance on unseen examples. Recently, large margin classification techniques emerged as a practical result of the theory of generalization, in particular Boosting and Support Vector Machines. A large margin implies a good generalization performance. Hence, we analyze how large the margins in boosting are and find an improved algorithm that is able to generate the maximum margin solution.<br /> <br /> o How can boosting methods be related to mathematical optimization techniques?<br /> To analyze the properties of the resulting classification or regression rule, it is of high importance to understand whether and under which conditions boosting converges. We show that boosting can be used to solve large scale constrained optimization problems, whose solutions are well characterizable. To show this, we relate boosting methods to methods known from mathematical optimization, and derive convergence guarantees for a quite general family of boosting algorithms.<br /> <br /> o How to make Boosting noise robust?<br /> One of the problems of current boosting techniques is that they are sensitive to noise in the training sample. In order to make boosting robust, we transfer the soft margin idea from support vector learning to boosting. We develop theoretically motivated regularized algorithms that exhibit a high noise robustness.<br /> <br /> o How to adapt boosting to regression problems?<br /> Boosting methods are originally designed for classification problems. To extend the boosting idea to regression problems, we use the previous convergence results and relations to semi-infinite programming to design boosting-like algorithms for regression problems. We show that these leveraging algorithms have desirable theoretical and practical properties.<br /> <br /> o Can boosting techniques be useful in practice?<br /> The presented theoretical results are guided by simulation results either to illustrate properties of the proposed algorithms or to show that they work well in practice. We report on successful applications in a non-intrusive power monitoring system, chaotic time series analysis and a drug discovery process. <br><br> ---<br> Anmerkung:<br> Der Autor ist Träger des von der Mathematisch-Naturwissenschaftlichen Fakultät der Universität Potsdam vergebenen Michelson-Preises für die beste Promotion des Jahres 2001/2002.

Brain neural activity generates electrical discharges, which manifest as electrical and magnetic potentials around the scalp. Those potentials can be registered with magnetoencephalography (MEG) and electroencephalography (EEG) devices. Data acquired by M/EEG is extremely difficult to work with due to the inherent complexity of underlying brain processes and low signal-to-noise ratio (SNR). Machine learning techniques have to be employed in order to reveal the underlying structure of the signal and to understand the brain state. This thesis explores a diverse range of machine learning techniques which model the structure of M/EEG data in order to decode the mental state. It focuses on measuring a subject's variability and on modeling intrasubject variability. We propose to measure subject variability with a spectral clustering setup. Further, we extend this approach to a unified classification framework based on Laplacian regularized support vector machine (SVM). We solve the issue of intrasubject variability by employing a model with latent variables (based on a latent SVM). Latent variables describe transformations that map samples into a comparable state. We focus mainly on intrasubject experiments to model temporal misalignment.

The main contribution of this thesis is the concept of Fenchel duality with a focus on its application in the field of machine learning problems and image restoration tasks. We formulate a general optimization problem for modeling support vector machine tasks and assign a Fenchel dual problem to it, prove weak and strong duality statements as well as necessary and sufficient optimality conditions for that primal-dual pair. In addition, several special instances of the general optimization problem are derived for different choices of loss functions for both the regression and the classifification task. The convenience of these approaches is demonstrated by numerically solving several problems. We formulate a general nonsmooth optimization problem and assign a Fenchel dual problem to it. It is shown that the optimal objective values of the primal and the dual one coincide and that the primal problem has an optimal solution under certain assumptions. The dual problem turns out to be nonsmooth in general and therefore a regularization is performed twice to obtain an approximate dual problem that can be solved efficiently via a fast gradient algorithm. We show how an approximate optimal and feasible primal solution can be constructed by means of some sequences of proximal points closely related to the dual iterates. Furthermore, we show that the solution will indeed converge to the optimal solution of the primal for arbitrarily small accuracy. Finally, the support vector regression task is obtained to arise as a particular case of the general optimization problem and the theory is specialized to this problem. We calculate several proximal points occurring when using difffferent loss functions as well as for some regularization problems applied in image restoration tasks. Numerical experiments illustrate the applicability of our approach for these types of problems.

This thesis examines the concept of image self-similarity and provides solutions to various associated inverse problems such as resolution enhancement and missing fractal codes. In general, many real-world inverse problems are ill-posed, mainly because of the lack of existence of a unique solution. The procedure of providing acceptable unique solutions to such problems is known as regularization. The concept of image prior, which has been of crucial importance in image modelling and processing, has also been important in solving inverse problems since it algebraically translates to the regularization procedure. Indeed, much recent progress in imaging has been due to advances in the formulation and practice of regularization. This, coupled with progress in optimization and numerical analysis, has yielded much improvement in computational methods of solving inverse imaging problems. Historically, the idea of self-similarity was important in the development of fractal image coding. Here we show that the self-similarity properties of natural images may be used to construct image priors for the purpose of addressing certain inverse problems. Indeed, new trends in the area of non-local image processing have provided a rejuvenated appreciation of image self-similarity and opportunities to explore novel self-similarity-based priors. We first revisit the concept of fractal-based methods and address some open theoretical problems in the area. This includes formulating a necessary and sufficient condition for the contractivity of the block fractal transform operator. We shall also provide some more generalized formulations of fractal-based self-similarity constraints of an image. These formulations can be developed algebraically and also in terms of the set-based method of Projection Onto Convex Sets (POCS). We then revisit the traditional inverse problems of single frame image zooming and multi-frame resolution enhancement, also known as super-resolution. Some ideas will be borrowed from newly developed non-local denoising algorithms in order to formulate self-similarity priors. Understanding the role of scale and choice of examples/samples is also important in these proposed models. For this purpose, we perform an extensive series of numerical experiments and analyze the results. These ideas naturally lead to the method of self-examples, which relies on the regularity properties of natural images at different scales, as a means of solving the single-frame image zooming problem. Furthermore, we propose and investigate a multi-frame super-resolution counterpart which does not require explicit motion estimation among video sequences.