Neuroimagiologia funcional é uma área da neurociência que visa o desenvolvimento de diversas técnicas para mapear a atividade do sistema nervoso e esteve sob constante desenvolvimento durante as últimas décadas devido à sua grande importância para aplicações clínicas e pesquisa. Técnicas usualmente utilizadas, como imagem por ressonância magnética functional (fMRI) e tomografia por emissão de pósitrons (PET) têm ótima resolução espacial (~ mm), mas uma resolução temporal limitada (~ s), impondo um grande desafio para nossa compreensão a respeito da dinâmica de funções cognitivas mais elevadas, cujas oscilações podem ocorrer em escalas temporais muito mais finas (~ ms). Tal limitação ocorre pelo fato destas técnicas medirem respostas biológicas lentas que são correlacionadas de maneira indireta com a atividade elétrica cerebral. As duas principais técnicas capazes de superar essa limitação são a Eletro- e Magnetoencefalografia (EEG/MEG), que são técnicas não invasivas para medir os campos elétricos e magnéticos no escalpo, respectivamente, gerados pelas fontes elétricas cerebrais. Ambas possuem resolução temporal na ordem de milisegundo, mas tipicalmente uma baixa resolução espacial (~ cm) devido à natureza mal posta do problema inverso eletromagnético. Um imenso esforço vem sendo feito durante as últimas décadas para melhorar suas resoluções espaciais através da incorporação de informação relevante ao problema de outras técnicas de imagens e/ou de vínculos biologicamente inspirados aliados ao desenvolvimento de métodos matemáticos e algoritmos sofisticados. Neste trabalho focaremos em EEG, embora todas técnicas aqui apresentadas possam ser igualmente aplicadas ao MEG devido às suas formas matemáticas idênticas. Em particular, nós exploramos esparsidade como uma importante restrição matemática dentro de uma abordagem Bayesiana chamada Aprendizagem Bayesiana Esparsa (SBL), que permite a obtenção de soluções únicas significativas no problema de reconstrução de fontes. Além disso, investigamos como incorporar diferentes estruturas como graus de liberdade nesta abordagem, que é uma aplicação de esparsidade estruturada e mostramos que é um caminho promisor para melhorar a precisão de reconstrução de fontes em métodos de imagens eletromagnéticos. / Functional Neuroimaging is an area of neuroscience which aims at developing several techniques to map the activity of the nervous system and has been under constant development in the last decades due to its high importance in clinical applications and research. Common applied techniques such as functional magnetic resonance imaging (fMRI) and positron emission tomography (PET) have great spatial resolution (~ mm), but a limited temporal resolution (~ s), which poses a great challenge on our understanding of the dynamics of higher cognitive functions, whose oscillations can occur in much finer temporal scales (~ ms). Such limitation occurs because these techniques rely on measurements of slow biological responses which are correlated in a complicated manner to the actual electric activity. The two major candidates that overcome this shortcoming are Electro- and Magnetoencephalography (EEG/MEG), which are non-invasive techniques that measure the electric and magnetic fields on the scalp, respectively, generated by the electrical brain sources. Both have millisecond temporal resolution, but typically low spatial resolution (~ cm) due to the highly ill-posed nature of the electromagnetic inverse problem. There has been a huge effort in the last decades to improve their spatial resolution by means of incorporating relevant information to the problem from either other imaging modalities and/or biologically inspired constraints allied with the development of sophisticated mathematical methods and algorithms. In this work we focus on EEG, although all techniques here presented can be equally applied to MEG because of their identical mathematical form. In particular, we explore sparsity as a useful mathematical constraint in a Bayesian framework called Sparse Bayesian Learning (SBL), which enables the achievement of meaningful unique solutions in the source reconstruction problem. Moreover, we investigate how to incorporate different structures as degrees of freedom into this framework, which is an application of structured sparsity and show that it is a promising way to improve the source reconstruction accuracy of electromagnetic imaging methods.

Neuroimagiologia funcional é uma área da neurociência que visa o desenvolvimento de diversas técnicas para mapear a atividade do sistema nervoso e esteve sob constante desenvolvimento durante as últimas décadas devido à sua grande importância para aplicações clínicas e pesquisa. Técnicas usualmente utilizadas, como imagem por ressonância magnética functional (fMRI) e tomografia por emissão de pósitrons (PET) têm ótima resolução espacial (~ mm), mas uma resolução temporal limitada (~ s), impondo um grande desafio para nossa compreensão a respeito da dinâmica de funções cognitivas mais elevadas, cujas oscilações podem ocorrer em escalas temporais muito mais finas (~ ms). Tal limitação ocorre pelo fato destas técnicas medirem respostas biológicas lentas que são correlacionadas de maneira indireta com a atividade elétrica cerebral. As duas principais técnicas capazes de superar essa limitação são a Eletro- e Magnetoencefalografia (EEG/MEG), que são técnicas não invasivas para medir os campos elétricos e magnéticos no escalpo, respectivamente, gerados pelas fontes elétricas cerebrais. Ambas possuem resolução temporal na ordem de milisegundo, mas tipicalmente uma baixa resolução espacial (~ cm) devido à natureza mal posta do problema inverso eletromagnético. Um imenso esforço vem sendo feito durante as últimas décadas para melhorar suas resoluções espaciais através da incorporação de informação relevante ao problema de outras técnicas de imagens e/ou de vínculos biologicamente inspirados aliados ao desenvolvimento de métodos matemáticos e algoritmos sofisticados. Neste trabalho focaremos em EEG, embora todas técnicas aqui apresentadas possam ser igualmente aplicadas ao MEG devido às suas formas matemáticas idênticas. Em particular, nós exploramos esparsidade como uma importante restrição matemática dentro de uma abordagem Bayesiana chamada Aprendizagem Bayesiana Esparsa (SBL), que permite a obtenção de soluções únicas significativas no problema de reconstrução de fontes. Além disso, investigamos como incorporar diferentes estruturas como graus de liberdade nesta abordagem, que é uma aplicação de esparsidade estruturada e mostramos que é um caminho promisor para melhorar a precisão de reconstrução de fontes em métodos de imagens eletromagnéticos. / Functional Neuroimaging is an area of neuroscience which aims at developing several techniques to map the activity of the nervous system and has been under constant development in the last decades due to its high importance in clinical applications and research. Common applied techniques such as functional magnetic resonance imaging (fMRI) and positron emission tomography (PET) have great spatial resolution (~ mm), but a limited temporal resolution (~ s), which poses a great challenge on our understanding of the dynamics of higher cognitive functions, whose oscillations can occur in much finer temporal scales (~ ms). Such limitation occurs because these techniques rely on measurements of slow biological responses which are correlated in a complicated manner to the actual electric activity. The two major candidates that overcome this shortcoming are Electro- and Magnetoencephalography (EEG/MEG), which are non-invasive techniques that measure the electric and magnetic fields on the scalp, respectively, generated by the electrical brain sources. Both have millisecond temporal resolution, but typically low spatial resolution (~ cm) due to the highly ill-posed nature of the electromagnetic inverse problem. There has been a huge effort in the last decades to improve their spatial resolution by means of incorporating relevant information to the problem from either other imaging modalities and/or biologically inspired constraints allied with the development of sophisticated mathematical methods and algorithms. In this work we focus on EEG, although all techniques here presented can be equally applied to MEG because of their identical mathematical form. In particular, we explore sparsity as a useful mathematical constraint in a Bayesian framework called Sparse Bayesian Learning (SBL), which enables the achievement of meaningful unique solutions in the source reconstruction problem. Moreover, we investigate how to incorporate different structures as degrees of freedom into this framework, which is an application of structured sparsity and show that it is a promising way to improve the source reconstruction accuracy of electromagnetic imaging methods.

L'imagerie par résonance magnétique (IRM) est la technique d'imagerie médicale de référence pour sonder in vivo et non invasivement les tissus mous du corps humain, en particulier le cerveau.L'amélioration de la résolution de l'IRM en un temps d'acquisition standard (400µm isotrope en 15 minutes) permettrait aux médecins d'améliorer considérablement leur diagnostic et le suivi des patients. Cependant, le temps d'acquisition en IRM reste long. Pour réduire ce temps, la récente théorie de l'échantillonnage comprimée (EC) a révolutionné la façon d'acquérir des données dans plusieurs domaines dont l'IRM en surmontant le théorème de Shannon-Nyquist. Avec l'EC, les données peuvent alors être massivement sous-échantillonnées tout en assurant des conditions optimales de reconstruction des images.Dans ce contexte, les thèses de doctorat précédemment soutenue au sein du laboratoire ont été consacrées à la conception et à la mise en oeuvre de scénarios d'acquisition physiquement plausibles pour accélérer l'acquisitions. Un nouvel algorithme d'optimisation pour la conception de trajectoire non cartésienne avancée appelée SPARKLING pour Spreading Projection Algorithm for Rapid K-space samplING en est né. Les trajectoires SPARKLING générées ont conduit à des facteurs d'accélération allant jusqu'à 20 en 2D et 70 pour les acquisitions 3D sur des images à haute résolution pondérées en T*₂ acquises à 7 Tesla. Ces accélérations n'étaient accessibles que grâce au rapport signal/bruit d'entrée élevé fourni par l'utilisation de bobines de réception multi-canaux (IRMp). Cependant, ces résultats ont été obtenus au détriment d'une reconstruction longue et complexe. Dans cette thèse, l'objectif est de proposer une nouvelle approche de reconstruction en ligne d'images acquies par IRMp non cartésiennes. Pour atteindre cet objectif, nous nous appuyons sur une approche en ligne où reconstruction et acquisition s'entremèlent. Par conséquent, la reconstruction débute avant la fin de l'acquisition et un résultat partiel est délivré au cours de l'examen. L'ensemble du pipeline est compatible avec une implémentation réelle à travers l'interface Gadgetron pour produire les images reconstruites à la console du scanner.Ainsi, après avoir exposé la théorie de l'échantillonage comprimé, nous présentons l'état de l'art de la méthode dédiée à la reconstruction en imagerie multi-canaux. En particulier, nous nous concentrerons d'abord sur les méthodes d'autocalibration qui présentent l'avantage d'être adaptées à l'échantillonnage non cartésien et nous proposons une méthode simple mais efficace pour estimer le profil de sensibilité des différents cannaux. Cependant, en raison de leur dépendance au profil de sensibilité, ces méthodes ne sont pas adapatable à la reconstruction en ligne. Par conséquent, la deuxième partie se concentre sur la suppression des ces profils et celà grâce à l'utilisation de norme mixte promouvant une parcimonie structurée. Ensuite, nous adaptons différentes réularization basées sur la parcimonie structurée pour reconstruire ces images fortement corrélées. Enfin, la méthode retenue sera appliquée à l'imagerie en ligne. / Magnetic resonance imaging (MRI) is the reference medical imaging technique for probing in vivo and non-invasively soft tissues in the human body, notably the brain. MR image resolution improvement in a standard scanning time (e.g., 400µm isotropic in 15 min) would allow medical doctors to significantly improve both their diagnosis and patients' follow-up. However the scanning time in MRI remains long, especially in the high resolution context. To reduce this time, the recent Compressed Sensing (CS) theory has revolutionized the way of acquiring data in several fields including MRI by overcoming the Shannon-Nyquist theorem. Using CS, data can then be massively under-sampled while ensuring conditions for optimal image recovery.In this context, previous Ph.D. thesis in the laboratory were dedicated to the design and implementation of physically plausible acquisition scenarios to accelerate the scan. Those projects deliver new optimization algorithm for the design of advanced non-Cartesian trajectory called SPARKLING: Spreading Projection Algorithm for Rapid K-space samplING. The generated SPARKLING trajectories led to acceleration factors up to 20 in 2D and 60 for 3D-acquisitions on highly resolved T₂* weighted images acquired at 7~Tesla.Those accelerations were only accessible thanks to the high input Signal-to-Noise Ratio delivered by the usage of multi-channel reception coils. However, those results are coming at a price of long and complex reconstruction.In this thesis, the objective is to propose an online approach for non-Cartesian multi-channel MR image reconstruction. To achieve this goal we rely on an online approach where the reconstruction starts from incomplete data.Hence acquisition and reconstruction are interleaved, and partial feedback is given during the scan. After exposing the Compressed Sensing theory, we present state-of the art method dedicated to multi-channel coil reconstruction. In particular, we will first focus on self-calibrating methods that presents the advantage to be adapted to non-Cartesian sampling and we propose a simple yet efficient method to estimate the coil sensitivity profile.However, owing to its dependence to user-defined parameters, this two-step approach (extraction of sensitivity maps and then image reconstruction) is not compatible with the timing constraints associated with online reconstruction. Then we studied the case of calibration-less reconstruction methods and splits them into two categories, the k-space based and the domain-based. While the k-space calibration-less method are sub-optimal for non-Cartesian reconstruction, due to the gridding procedure, we will retain the domain-based calibration-less reconstruction and prove theirs for online purposes. Hence in the second part, we first prove the advantage of mixed norm to improve the recovery guarantee in the pMRI setting. Then we studied the impact of structured sparse induced norm on the reconstruction multi-channel purposes, where then and adapt different penalty based on structured sparsity to handle those highly correlated images. Finally, the retained method will be applied to online purposes. The entire pipeline, is compatible with an implementation through the Gadgetron pipeline to deliver the reconstruction at the scanner console.

Visual object category recognition is one of the most challenging problems in computer vision. Even assuming that we can obtain a near-perfect instance level representation with the advances in visual input devices and low-level vision techniques, object categorization still remains as a difficult problem because it requires drawing boundaries between instances in a continuous world, where the boundaries are solely defined by human conceptualization. Object categorization is essentially a perceptual process that takes place in a human-defined semantic space. In this semantic space, the categories reside not in isolation, but in relation to others. Some categories are similar, grouped, or co-occur, and some are not. However, despite this semantic nature of object categorization, most of the today's automatic visual category recognition systems rely only on the category labels for training discriminative recognition with statistical machine learning techniques. In many cases, this could result in the recognition model being misled into learning incorrect associations between visual features and the semantic labels, from essentially overfitting to training set biases. This limits the model's prediction power when new test instances are given. Using semantic knowledge has great potential to benefit object category recognition. First, semantic knowledge could guide the training model to learn a correct association between visual features and the categories. Second, semantics provide much richer information beyond the membership information given by the labels, in the form of inter-category and category-attribute distances, relations, and structures. Finally, the semantic knowledge scales well as the relations between categories become larger with an increasing number of categories. My goal in this thesis is to learn discriminative models for categorization that leverage semantic knowledge for object recognition, with a special focus on the semantic relationships among different categories and concepts. To this end, I explore three semantic sources, namely attributes, taxonomies, and analogies, and I show how to incorporate them into the original discriminative model as a form of structural regularization. In particular, for each form of semantic knowledge I present a feature learning approach that defines a semantic embedding to support the object categorization task. The regularization penalizes the models that deviate from the known structures according to the semantic knowledge provided. The first semantic source I explore is attributes, which are human-describable semantic characteristics of an instance. While the existing work treated them as mid-level features which did not introduce new information, I focus on their potential as a means to better guide the learning of object categories, by enforcing the object category classifiers to share features with attribute classifiers, in a multitask feature learning framework. This approach essentially discovers the common low-dimensional features that support predictions in both semantic spaces. Then, I move on to the semantic taxonomy, which is another valuable source of semantic knowledge. The merging and splitting criteria for the categories on a taxonomy are human-defined, and I aim to exploit this implicit semantic knowledge. Specifically, I propose a tree of metrics (ToM) that learns metrics that capture granularity-specific similarities at different nodes of a given semantic taxonomy, and uses a regularizer to isolate granularity-specific disjoint features. This approach captures the intuition that the features used for the discrimination of the parent class should be different from the features used for the children classes. Such learned metrics can be used for hierarchical classification. The use of a single taxonomy can be limited in that its structure is not optimal for hierarchical classification, and there may exist no single optimal semantic taxonomy that perfectly aligns with visual distributions. Thus, I next propose a way to overcome this limitation by leveraging multiple taxonomies as semantic sources to exploit, and combine the acquired complementary information across multiple semantic views and granularities. This allows us, for example, to synthesize semantics from both 'Biological', and 'Appearance'-based taxonomies when learning the visual features. Finally, as a further exploration of more complex semantic relations different from the previous two pairwise similarity-based models, I exploit analogies, which encode the relational similarities between two related pairs of categories. Specifically, I use analogies to regularize a discriminatively learned semantic embedding space for categorization, such that the displacements between the two category embeddings in both category pairs of the analogy are enforced to be the same. Such a constraint allows for a more confusing pair of categories to benefit from a clear separation in the matched pair of categories that share the same relation. All of these methods are evaluated on challenging public datasets, and are shown to effectively improve the recognition accuracy over purely discriminative models, while also guiding the recognition to be more semantic to human perception. Further, the applications of the proposed methods are not limited to visual object categorization in computer vision, but they can be applied to any classification problems where there exists some domain knowledge about the relationships or structures between the classes. Possible applications of my methods outside the visual recognition domain include document classification in natural language processing, and gene-based animal or protein classification in computational biology. / text

abstract: Deep neural networks (DNN) have shown tremendous success in various cognitive tasks, such as image classification, speech recognition, etc. However, their usage on resource-constrained edge devices has been limited due to high computation and large memory requirement. To overcome these challenges, recent works have extensively investigated model compression techniques such as element-wise sparsity, structured sparsity and quantization. While most of these works have applied these compression techniques in isolation, there have been very few studies on application of quantization and structured sparsity together on a DNN model. This thesis co-optimizes structured sparsity and quantization constraints on DNN models during training. Specifically, it obtains optimal setting of 2-bit weight and 2-bit activation coupled with 4X structured compression by performing combined exploration of quantization and structured compression settings. The optimal DNN model achieves 50X weight memory reduction compared to floating-point uncompressed DNN. This memory saving is significant since applying only structured sparsity constraints achieves 2X memory savings and only quantization constraints achieves 16X memory savings. The algorithm has been validated on both high and low capacity DNNs and on wide-sparse and deep-sparse DNN models. Experiments demonstrated that deep-sparse DNN outperforms shallow-dense DNN with varying level of memory savings depending on DNN precision and sparsity levels. This work further proposed a Pareto-optimal approach to systematically extract optimal DNN models from a huge set of sparse and dense DNN models. The resulting 11 optimal designs were further evaluated by considering overall DNN memory which includes activation memory and weight memory. It was found that there is only a small change in the memory footprint of the optimal designs corresponding to the low sparsity DNNs. However, activation memory cannot be ignored for high sparsity DNNs. / Dissertation/Thesis / Masters Thesis Computer Engineering 2018

This thesis deals with audio inpainting problem. Firstly, basic concepts are summarized. Then, sparse representation of signals is introduced along with several algorithms. In the main part dedicated to the audio inpainting, the problem is defined and actual methods are presented and compared. The newest approach using the harmonic strucure of sound signals is then introduced, followed by several experiments and evaluation. Lastly, an algorithm ensuring the maximal computational efficiency is derived.

Le traitement massif et automatique des données requiert le développement de techniques de filtration des informations les plus importantes. Parmi ces méthodes, celles présentant des structures parcimonieuses se sont révélées idoines pour améliorer l’efficacité statistique et computationnelle des estimateurs, dans un contexte de grandes dimensions. Elles s’expriment souvent comme solution de la minimisation du risque empirique régularisé s’écrivant comme une somme d’un terme lisse qui mesure la qualité de l’ajustement aux données, et d’un terme non lisse qui pénalise les solutions complexes. Cependant, une telle manière d’inclure des informations a priori, introduit de nombreuses difficultés numériques pour résoudre le problème d’optimisation sous-jacent et pour calibrer le niveau de régularisation. Ces problématiques ont été au coeur des questions que nous avons abordées dans cette thèse.Une technique récente, appelée «Screening Rules», propose d’ignorer certaines variables pendant le processus d’optimisation en tirant bénéfice de la parcimonie attendue des solutions. Ces règles d’élimination sont dites sûres lorsqu’elles garantissent de ne pas rejeter les variables à tort. Nous proposons un cadre unifié pour identifier les structures importantes dans ces problèmes d’optimisation convexes et nous introduisons les règles «Gap Safe Screening Rules». Elles permettent d’obtenir des gains considérables en temps de calcul grâce à la réduction de la dimension induite par cette méthode. De plus, elles s’incorporent facilement aux algorithmes itératifs et s’appliquent à un plus grand nombre de problèmes que les méthodes précédentes.Pour trouver un bon compromis entre minimisation du risque et introduction d’un biais d’apprentissage, les algorithmes d’homotopie offrent la possibilité de tracer la courbe des solutions en fonction du paramètre de régularisation. Toutefois, ils présentent des instabilités numériques dues à plusieurs inversions de matrice, et sont souvent coûteux en grande dimension. Aussi, ils ont des complexités exponentielles en la dimension du modèle dans des cas défavorables. En autorisant des solutions approchées, une approximation de la courbe des solutions permet de contourner les inconvénients susmentionnés. Nous revisitons les techniques d’approximation des chemins de régularisation pour une tolérance prédéfinie, et nous analysons leur complexité en fonction de la régularité des fonctions de perte en jeu. Il s’ensuit une proposition d’algorithmes optimaux ainsi que diverses stratégies d’exploration de l’espace des paramètres. Ceci permet de proposer une méthode de calibration de la régularisation avec une garantie de convergence globale pour la minimisation du risque empirique sur les données de validation.Le Lasso, un des estimateurs parcimonieux les plus célèbres et les plus étudiés, repose sur une théorie statistique qui suggère de choisir la régularisation en fonction de la variance des observations. Ceci est difficilement utilisable en pratique car, la variance du modèle est une quantité souvent inconnue. Dans de tels cas, il est possible d’optimiser conjointement les coefficients de régression et le niveau de bruit. Ces estimations concomitantes, apparues dans la littérature sous les noms de Scaled Lasso, Square-Root Lasso, fournissent des résultats théoriques aussi satisfaisants que celui du Lasso tout en étant indépendant de la variance réelle. Bien que présentant des avancées théoriques et pratiques importantes, ces méthodes sont aussi numériquement instables et les algorithmes actuellement disponibles sont coûteux en temps de calcul. Nous illustrons ces difficultés et nous proposons à la fois des modifications basées sur des techniques de lissage pour accroitre la stabilité numérique de ces estimateurs, ainsi qu’un algorithme plus efficace pour les obtenir. / Massive and automatic data processing requires the development of techniques able to filter the most important information. Among these methods, those with sparse structures have been shown to improve the statistical and computational efficiency of estimators in a context of large dimension. They can often be expressed as a solution of regularized empirical risk minimization and generally lead to non differentiable optimization problems in the form of a sum of a smooth term, measuring the quality of the fit, and a non-smooth term, penalizing complex solutions. Although it has considerable advantages, such a way of including prior information, unfortunately introduces many numerical difficulties both for solving the underlying optimization problem and to calibrate the level of regularization. Solving these issues has been at the heart of this thesis. A recently introduced technique, called "Screening Rules", proposes to ignore some variables during the optimization process by benefiting from the expected sparsity of the solutions. These elimination rules are said to be safe when the procedure guarantees to not reject any variable wrongly. In this work, we propose a unified framework for identifying important structures in these convex optimization problems and we introduce the "Gap Safe Screening Rules". They allows to obtain significant gains in computational time thanks to the dimensionality reduction induced by this method. In addition, they can be easily inserted into iterative algorithms and apply to a large number of problems.To find a good compromise between minimizing risk and introducing a learning bias, (exact) homotopy continuation algorithms offer the possibility of tracking the curve of the solutions as a function of the regularization parameters. However, they exhibit numerical instabilities due to several matrix inversions and are often expensive in large dimension. Another weakness is that a worst-case analysis shows that they have exact complexities that are exponential in the dimension of the model parameter. Allowing approximated solutions makes possible to circumvent the aforementioned drawbacks by approximating the curve of the solutions. In this thesis, we revisit the approximation techniques of the regularization paths given a predefined tolerance and we propose an in-depth analysis of their complexity w.r.t. the regularity of the loss functions involved. Hence, we propose optimal algorithms as well as various strategies for exploring the parameters space. We also provide calibration method (for the regularization parameter) that enjoys globalconvergence guarantees for the minimization of the empirical risk on the validation data.Among sparse regularization methods, the Lasso is one of the most celebrated and studied. Its statistical theory suggests choosing the level of regularization according to the amount of variance in the observations, which is difficult to use in practice because the variance of the model is oftenan unknown quantity. In such case, it is possible to jointly optimize the regression parameter as well as the level of noise. These concomitant estimates, appeared in the literature under the names of Scaled Lasso or Square-Root Lasso, and provide theoretical results as sharp as that of theLasso while being independent of the actual noise level of the observations. Although presenting important advances, these methods are numerically unstable and the currently available algorithms are expensive in computation time. We illustrate these difficulties and we propose modifications based on smoothing techniques to increase stability of these estimators as well as to introduce a faster algorithm.

L’imagerie optique par colorant potentiométrique est une méthode d’enregistrement de l’activité corticale prometteuse, mais dont le potentiel réel est limité par la présence d’artefacts et d’interférences dans les acquisitions. À partir de modèles existant dans la littérature, nous proposons un modèle génératif du signal basé sur un mélange additif de composantes, chacune contrainte dans une union d’espaces linéaires déterminés par son origine biophysique. Motivés par le problème de séparation de composantes qui en découle, qui est un problème inverse linéaire sous-déterminé, nous développons : (1) des régularisations convexes structurées spatialement, favorisant en particulier des solutions parcimonieuses ; (2) un nouvel algorithme proximal de premier ordre pour minimiser efficacement la fonctionnelle qui en résulte ; (3) des méthodes statistiques de sélection de paramètre basées sur l’estimateur non biaisé du risque de Stein. Nous étudions ces outils dans un cadre général, et discutons leur utilité pour de nombreux domaines des mathématiques appliqués, en particulier pour les problèmes inverses ou de régression en grande dimension. Nous développons par la suite un logiciel de séparation de composantes en présence de bruit, dans un environnement intégré adapté à l’imagerie optique par colorant potentiométrique. Finalement, nous évaluons ce logiciel sur différentes données, synthétiques et réelles, montrant des résultats encourageants quant à la possibilité d’observer des dynamiques corticales complexes. / Voltage-sensitive dye optical imaging is a promising recording modality for the cortical activity, but its practical potential is limited by many artefacts and interferences in the acquisitions. Inspired by existing models in the literature, we propose a generative model of the signal, based on an additive mixtures of components, each one being constrained within an union of linear spaces, determined by its biophysical origin. Motivated by the resulting component separation problem, which is an underdetermined linear inverse problem, we develop: (1) convex, spatially structured regularizations, enforcing in particular sparsity on the solutions; (2) a new rst-order proximal algorithm for minimizing e›ciently the resulting functional; (3) statistical methods for automatic parameters selection, based on Stein’s unbiased risk estimate.We study thosemethods in a general framework, and discuss their potential applications in variouselds of applied mathematics, in particular for large scale inverse problems or regressions. We develop subsequently a soŸware for noisy component separation, in an integrated environment adapted to voltage-sensitive dye optical imaging. Finally, we evaluate this soŸware on dišerent data set, including synthetic and real data, showing encouraging perspectives for the observation of complex cortical dynamics.

No description available.

Les algorithmes convexes de résolution pour les régressions linéaires parcimonieuses possèdent de bonnes performances pratiques et théoriques. Cependant, ils souffrent tous des dimensions du problème qui dictent la complexité de chacune de leur itération. Nous proposons une approche pour réduire ce coût calculatoire au niveau de l'itération. Des stratégies récentes s'appuyant sur des tests d'élimination de variables ont été proposées pour accélérer la résolution des problèmes de régressions parcimonieuse pénalisées tels que le LASSO. Ces approches reposent sur l'idée qu'il est profitable de dédier un petit effort de calcul pour localiser des atomes inactifs afin de les retirer du dictionnaire dans une étape de prétraitement. L'algorithme de résolution utilisant le dictionnaire ainsi réduit convergera alors plus rapidement vers la solution du problème initial. Nous pensons qu'il existe un moyen plus efficace pour réduire le dictionnaire et donc obtenir une meilleure accélération : à l'intérieur de chaque itération de l'algorithme, il est possible de valoriser les calculs originalement dédiés à l'algorithme pour obtenir à moindre coût un nouveau test d'élimination dont l'effet d'élimination augmente progressivement le long des itérations. Le dictionnaire est alors réduit de façon dynamique au lieu d'être réduit de façon statique, une fois pour toutes, avant la première itération. Nous formalisons ce principe d'élimination dynamique à travers une formulation algorithmique générique, et l'appliquons en intégrant des tests d'élimination existants, à l'intérieur de plusieurs algorithmes du premier ordre pour résoudre les problèmes du LASSO et Group-LASSO. / Applications in signal processing and machine learning make frequent use of sparse regressions. Resulting convex problems, such as the LASSO, can be efficiently solved thanks to first-order algorithms, which are general, and have good convergence properties. However those algorithms suffer from the dimension of the problem, which impose the complexity of their iterations. In this thesis we study approaches, based on screening tests, aimed at reducing the computational cost at the iteration level. Such approaches build upon the idea that it is worth dedicating some small computational effort to locate inactive atoms and remove them from the dictionary in a preprocessing stage so that the regression algorithm working with a smaller dictionary will then converge faster to the solution of the initial problem. We believe that there is an even more efficient way to screen the dictionary and obtain a greater acceleration: inside each iteration of the regression algorithm, one may take advantage of the algorithm computations to obtain a new screening test for free with increasing screening effects along the iterations. The dictionary is henceforth dynamically screened instead of being screened statically, once and for all, before the first iteration. Our first contribution is the formalisation of this principle and its application to first-order algorithms, for the resolution of the LASSO and Group-LASSO. In a second contribution, this general principle is combined to active-set methods, whose goal is also to accelerate the resolution of sparse regressions. Applying the two complementary methods on first-order algorithms, leads to great acceleration performances.