Ce travail de thèse porte sur l'obtention de bornes de généralisation pour des échantillons statistiques à valeur dans des espaces de Hilbert définis par des noyaux reproduisants. L'approche consiste à obtenir des bornes non asymptotiques indépendantes de la dimension dans des espaces de dimension finie, en utilisant des inégalités PAC-Bayesiennes liées à une perturbation Gaussienne du paramètre et à les étendre ensuite aux espaces de Hilbert séparables. On se pose dans un premier temps la question de l'estimation de l'opérateur de Gram à partir d'un échantillon i. i. d. par un estimateur robuste et on propose des bornes uniformes, sous des hypothèses faibles de moments. Ces résultats permettent de caractériser l'analyse en composantes principales indépendamment de la dimension et d'en proposer des variantes robustes. On propose ensuite un nouvel algorithme de clustering spectral. Au lieu de ne garder que la projection sur les premiers vecteurs propres, on calcule une itérée du Laplacian normalisé. Cette itération, justifiée par l'analyse du clustering en termes de chaînes de Markov, opère comme une version régularisée de la projection sur les premiers vecteurs propres et permet d'obtenir un algorithme dans lequel le nombre de clusters est déterminé automatiquement. On présente des bornes non asymptotiques concernant la convergence de cet algorithme, lorsque les points à classer forment un échantillon i. i. d. d'une loi à support compact dans un espace de Hilbert. Ces bornes sont déduites des bornes obtenues pour l'estimation d'un opérateur de Gram dans un espace de Hilbert. On termine par un aperçu de l'intérêt du clustering spectral dans le cadre de l'analyse d'images. / This thesis focuses on obtaining generalization bounds for random samples in reproducing kernel Hilbert spaces. The approach consists in first obtaining non-asymptotic dimension-free bounds in finite-dimensional spaces using some PAC-Bayesian inequalities related to Gaussian perturbations and then in generalizing the results in a separable Hilbert space. We first investigate the question of estimating the Gram operator by a robust estimator from an i. i. d. sample and we present uniform bounds that hold under weak moment assumptions. These results allow us to qualify principal component analysis independently of the dimension of the ambient space and to propose stable versions of it. In the last part of the thesis we present a new algorithm for spectral clustering. It consists in replacing the projection on the eigenvectors associated with the largest eigenvalues of the Laplacian matrix by a power of the normalized Laplacian. This iteration, justified by the analysis of clustering in terms of Markov chains, performs a smooth truncation. We prove nonasymptotic bounds for the convergence of our spectral clustering algorithm applied to a random sample of points in a Hilbert space that are deduced from the bounds for the Gram operator in a Hilbert space. Experiments are done in the context of image analysis.

In the past few years, complex neural networks have achieved state of the art results in image classification. However, training these models requires large amounts of labelled data. Whereas unlabelled images are often readily available in large quantities, obtaining l abels takes considerable human effort. Active learning reduces the required labelling effort by selecting the most informative instances to label. The most popular active learning query strategy framework, uncertainty sampling, uses uncertainty estimates of the model being trained to select instances for labelling. However, modern classification neural networks often do not provide good uncertainty estimates. Baye sian neural networks model uncertainties over model parameters, which can be used to obtain uncertainties over model predictions. Exact Bayesian inference is intractable for neural networks, however several approximate methods have been proposed. We experiment with three such methods using various uncertainty sampling active learning query strategies.

<div>Using sensor measurements to estimate the states and parameters of a system is a fundamental task in understanding the behavior of the system. Moreover, as modern systems grow rapidly in scale and complexity, it is not always possible to deploy sensors to measure all of the states and parameters of the system, due to cost and physical constraints. Therefore, selecting an optimal subset of all the candidate sensors to deploy and gather measurements of the system is an important and challenging problem. In addition, the systems may be targeted by external attackers who attempt to remove or destroy the deployed sensors. This further motivates the formulation of resilient sensor selection strategies. In this thesis, we address the sensor selection problem under different settings as follows. </div><div><br></div><div>First, we consider the optimal sensor selection problem for linear dynamical systems with stochastic inputs, where the Kalman filter is applied based on the sensor measurements to give an estimate of the system states. The goal is to select a subset of sensors under certain budget constraints such that the trace of the steady-state error covariance of the Kalman filter with the selected sensors is minimized. We characterize the complexity of this problem by showing that the Kalman filtering sensor selection problem is NP-hard and cannot be approximated within any constant factor in polynomial time for general systems. We then consider the optimal sensor attack problem for Kalman filtering. The Kalman filtering sensor attack problem is to attack a subset of selected sensors under certain budget constraints in order to maximize the trace of the steady-state error covariance of the Kalman filter with sensors after the attack. We show that the same results as the Kalman filtering sensor selection problem also hold for the Kalman filtering sensor attack problem. Having shown that the general sensor selection and sensor attack problems for Kalman filtering are hard to solve, our next step is to consider special classes of the general problems. Specifically, we consider the underlying directed network corresponding to a linear dynamical system and investigate the case when there is a single node of the network that is affected by a stochastic input. In this setting, we show that the corresponding sensor selection and sensor attack problems for Kalman filtering can be solved in polynomial time. We further study the resilient sensor selection problem for Kalman filtering, where the problem is to find a sensor selection strategy under sensor selection budget constraints such that the trace of the steady-state error covariance of the Kalman filter is minimized after an adversary removes some of the deployed sensors. We show that the resilient sensor selection problem for Kalman filtering is NP-hard, and provide a pseudo-polynomial-time algorithm to solve it optimally.</div><div> </div><div> Next, we consider the sensor selection problem for binary hypothesis testing. The problem is to select a subset of sensors under certain budget constraints such that a certain metric of the Neyman-Pearson (resp., Bayesian) detector corresponding to the selected sensors is optimized. We show that this problem is NP-hard if the objective is to minimize the miss probability (resp., error probability) of the Neyman-Pearson (resp., Bayesian) detector. We then consider three optimization objectives based on the Kullback-Leibler distance, J-Divergence and Bhattacharyya distance, respectively, in the hypothesis testing sensor selection problem, and provide performance bounds on greedy algorithms when applied to the sensor selection problem associated with these optimization objectives.</div><div> </div><div> Moving beyond the binary hypothesis setting, we also consider the setting where the true state of the world comes from a set that can have cardinality greater than two. A Bayesian approach is then used to learn the true state of the world based on the data streams provided by the data sources. We formulate the Bayesian learning data source selection problem under this setting, where the goal is to minimize the cost spent on the data sources such that the learning error is within a certain range. We show that the Bayesian learning data source selection is also NP-hard, and provide greedy algorithms with performance guarantees.</div><div> </div><div> Finally, in light of the COVID-19 pandemic, we study the parameter estimation measurement selection problem for epidemics spreading in networks. Here, the measurements (with certain costs) are collected by conducting virus and antibody tests on the individuals in the epidemic spread network. The goal of the problem is then to optimally estimate the parameters (i.e., the infection rate and the recovery rate of the virus) in the epidemic spread network, while satisfying the budget constraint on collecting the measurements. Again, we show that the measurement selection problem is NP-hard, and provide approximation algorithms with performance guarantees.</div>

No description available.

Many have used the principles of statistics and Bayesian decision theory to model specific learning problems. It is less common to see models of the processes of learning in general. One exception is the model of the supervised learning process known as the "Extended Bayesian Formalism" or EBF. This model is descriptive, in that it can describe and compare learning algorithms. Thus the EBF is capable of modeling both effective and ineffective learning algorithms. We extend the EBF to model un-supervised learning, semi-supervised learning, supervised learning, and empirical function optimization. We also generalize the utility model of the EBF to deal with non-deterministic outcomes, and with utility functions other than 0-1 loss. Finally, we modify the EBF to create a "prescriptive" learning model, meaning that, instead of describing existing algorithms, our model defines how learning should optimally take place. We call the resulting model the Unified Bayesian Decision Theoretical Model, or the UBDTM. WE show that this model can serve as a cohesive theory and framework in which a broad range of questions can be analyzed and studied. Such a broadly applicable unified theoretical framework is one of the major missing ingredients of machine learning theory. Using the UBDTM, we concentrate on supervised learning and empirical function optimization. We then use the UBDTM to reanalyze many important theoretical issues in Machine Learning, including No-Free-Lunch, utility implications, and active learning. We also point forward to future directions for using the UBDTM to model learnability, sample complexity, and ensembles. We also provide practical applications of the UBDTM by using the model to train a Bayesian variation to the CMAC supervised learner in closed form, to perform a practical empirical function optimization task, and as part of the guiding principles behind an ongoing project to create an electronic and print corpus of tagged ancient Syriac texts using active learning.