<div>Structured prediction consists of receiving a structured input and producing a combinatorial structure such as trees, clusters, networks, sequences, permutations, among others. From the computational viewpoint, structured prediction is in general considered <i>intractable</i> because of the size of the output space being exponential in the input size. For instance, in image segmentation tasks, the number of admissible segments is exponential in the number of pixels. A second factor is the combination of the input dimensionality along with the amount of data under availability. In structured prediction it is common to have the input live in a high-dimensional space, which involves to jointly reason about thousands or millions of variables, and at the same time contend with limited amount of data. Thus, learning and inference methods with strong computational and statistical guarantees are desired. The focus of our research is then to propose <i>principled methods</i> for structured prediction that are both polynomial time, i.e., <i>computationally efficient</i>, and require a polynomial number of data samples, i.e., <i>statistically efficient</i>.</div><div><br></div><div>The main contributions of this thesis are as follows:</div><div><br></div><div>(i) We develop an efficient and principled learning method of latent variable models for structured prediction under Gaussian perturbations. We derive a Rademacher-based generalization bound and argue that the use of non-convex formulations in learning latent-variable models leads to tighter bounds of the Gibbs decoder distortion.</div><div><br></div><div>(ii) We study the fundamental limits of structured prediction, i.e., we characterize the necessary sample complexity for learning factor graph models in the context of structured prediction. In particular, we show that the finiteness of our novel MaxPair-dimension is necessary for learning. Lastly, we show a connection between the MaxPair-dimension and the VC-dimension---which allows for using existing results on VC-dimension to calculate the MaxPair-dimension.</div><div><br></div><div>(iii) We analyze a generative model based on connected graphs, and find the structural conditions of the graph that allow for the exact recovery of the node labels. In particular, we show that exact recovery is realizable in polynomial time for a large class of graphs. Our analysis is based on convex relaxations, where we thoroughly analyze a semidefinite program and a degree-4 sum-of-squares program. Finally, we extend this model to consider linear constraints (e.g., fairness), and formally explain the effect of the added constraints on the probability of exact recovery.</div><div><br></div>

Ces dernières années, l'importance cruciale des métriques en apprentissage automatique a mené à un intérêt grandissant pour l'optimisation de distances et de similarités en utilisant l'information contenue dans des données d'apprentissage pour les rendre adaptées au problème traité. Ce domaine de recherche est souvent appelé apprentissage de métriques. En général, les méthodes existantes optimisent les paramètres d'une métrique devant respecter des contraintes locales sur les données d'apprentissage. Les métriques ainsi apprises sont généralement utilisées dans des algorithmes de plus proches voisins ou de clustering.Concernant les données numériques, beaucoup de travaux ont porté sur l'apprentissage de distance de Mahalanobis, paramétrisée par une matrice positive semi-définie. Les méthodes récentes sont capables de traiter des jeux de données de grande taille.Moins de travaux ont été dédiés à l'apprentissage de métriques pour les données structurées (comme les chaînes ou les arbres), car cela implique souvent des procédures plus complexes. La plupart des travaux portent sur l'optimisation d'une notion de distance d'édition, qui mesure (en termes de nombre d'opérations) le coût de transformer un objet en un autre.Au regard de l'état de l'art, nous avons identifié deux limites importantes des approches actuelles. Premièrement, elles permettent d'améliorer la performance d'algorithmes locaux comme les k plus proches voisins, mais l'apprentissage de métriques pour des algorithmes globaux (comme les classifieurs linéaires) n'a pour l'instant pas été beaucoup étudié. Le deuxième point, sans doute le plus important, est que la question de la capacité de généralisation des méthodes d'apprentissage de métriques a été largement ignorée.Dans cette thèse, nous proposons des contributions théoriques et algorithmiques qui répondent à ces limites. Notre première contribution est la construction d'un nouveau noyau construit à partir de probabilités d'édition apprises. A l'inverse d'autres noyaux entre chaînes, sa validité est garantie et il ne comporte aucun paramètre. Notre deuxième contribution est une nouvelle approche d'apprentissage de similarités d'édition pour les chaînes et les arbres inspirée par la théorie des (epsilon,gamma,tau)-bonnes fonctions de similarité et formulée comme un problème d'optimisation convexe. En utilisant la notion de stabilité uniforme, nous établissons des garanties théoriques pour la similarité apprise qui donne une borne sur l'erreur en généralisation d'un classifieur linéaire construit à partir de cette similarité. Dans notre troisième contribution, nous étendons ces principes à l'apprentissage de métriques pour les données numériques en proposant une méthode d'apprentissage de similarité bilinéaire qui optimise efficacement l'(epsilon,gamma,tau)-goodness. La similarité est apprise sous contraintes globales, plus appropriées à la classification linéaire. Nous dérivons des garanties théoriques pour notre approche, qui donnent de meilleurs bornes en généralisation pour le classifieur que dans le cas des données structurées. Notre dernière contribution est un cadre théorique permettant d'établir des bornes en généralisation pour de nombreuses méthodes existantes d'apprentissage de métriques. Ce cadre est basé sur la notion de robustesse algorithmique et permet la dérivation de bornes pour des fonctions de perte et des régulariseurs variés / In recent years, the crucial importance of metrics in machine learningalgorithms has led to an increasing interest in optimizing distanceand similarity functions using knowledge from training data to make them suitable for the problem at hand.This area of research is known as metric learning. Existing methods typically aim at optimizing the parameters of a given metric with respect to some local constraints over the training sample. The learned metrics are generally used in nearest-neighbor and clustering algorithms.When data consist of feature vectors, a large body of work has focused on learning a Mahalanobis distance, which is parameterized by a positive semi-definite matrix. Recent methods offer good scalability to large datasets.Less work has been devoted to metric learning from structured objects (such as strings or trees), because it often involves complex procedures. Most of the work has focused on optimizing a notion of edit distance, which measures (in terms of number of operations) the cost of turning an object into another.We identify two important limitations of current supervised metric learning approaches. First, they allow to improve the performance of local algorithms such as k-nearest neighbors, but metric learning for global algorithms (such as linear classifiers) has not really been studied so far. Second, and perhaps more importantly, the question of the generalization ability of metric learning methods has been largely ignored.In this thesis, we propose theoretical and algorithmic contributions that address these limitations. Our first contribution is the derivation of a new kernel function built from learned edit probabilities. Unlike other string kernels, it is guaranteed to be valid and parameter-free. Our second contribution is a novel framework for learning string and tree edit similarities inspired by the recent theory of (epsilon,gamma,tau)-good similarity functions and formulated as a convex optimization problem. Using uniform stability arguments, we establish theoretical guarantees for the learned similarity that give a bound on the generalization error of a linear classifier built from that similarity. In our third contribution, we extend the same ideas to metric learning from feature vectors by proposing a bilinear similarity learning method that efficiently optimizes the (epsilon,gamma,tau)-goodness. The similarity is learned based on global constraints that are more appropriate to linear classification. Generalization guarantees are derived for our approach, highlighting that our method minimizes a tighter bound on the generalization error of the classifier. Our last contribution is a framework for establishing generalization bounds for a large class of existing metric learning algorithms. It is based on a simple adaptation of the notion of algorithmic robustness and allows the derivation of bounds for various loss functions and regularizers.

This thesis studies the generalization ability of machine learning algorithms in a statistical setting. It focuses on the data-dependent analysis of the generalization performance of learning algorithms in order to make full use of the potential of the actual training sample from which these algorithms learn.¶ First, we propose an extension of the standard framework for the derivation of generalization bounds for algorithms taking their hypotheses from random classes of functions. This approach is motivated by the fact that the function produced by a learning algorithm based on a random sample of data depends on this sample and is therefore a random function. Such an approach avoids the detour of the worst-case uniform bounds as done in the standard approach. We show that the mechanism which allows one to obtain generalization bounds for random classes in our framework is based on a “small complexity” of certain random coordinate projections. We demonstrate how this notion of complexity relates to learnability and how one can explore geometric properties of these projections in order to derive estimates of rates of convergence and good confidence interval estimates for the expected risk. We then demonstrate the generality of our new approach by presenting a range of examples, among them the algorithm-dependent compression schemes and the data-dependent luckiness frameworks, which fall into our random subclass framework.¶ Second, we study in more detail generalization bounds for a specific algorithm which is of central importance in learning theory, namely the Empirical Risk Minimization algorithm (ERM). Recent results show that one can significantly improve the high-probability estimates for the convergence rates for empirical minimizers by a direct analysis of the ERM algorithm. These results are based on a new localized notion of complexity of subsets of hypothesis functions with identical expected errors and are therefore dependent on the underlying unknown distribution. We investigate the extent to which one can estimate these high-probability convergence rates in a data-dependent manner. We provide an algorithm which computes a data-dependent upper bound for the expected error of empirical minimizers in terms of the “complexity” of data-dependent local subsets. These subsets are sets of functions of empirical errors of a given range and can be determined based solely on empirical data. We then show that recent direct estimates, which are essentially sharp estimates on the high-probability convergence rate for the ERM algorithm, can not be recovered universally from empirical data.

In recent years, the crucial importance of metrics in machine learningalgorithms has led to an increasing interest in optimizing distanceand similarity functions using knowledge from training data to make them suitable for the problem at hand.This area of research is known as metric learning. Existing methods typically aim at optimizing the parameters of a given metric with respect to some local constraints over the training sample. The learned metrics are generally used in nearest-neighbor and clustering algorithms.When data consist of feature vectors, a large body of work has focused on learning a Mahalanobis distance, which is parameterized by a positive semi-definite matrix. Recent methods offer good scalability to large datasets.Less work has been devoted to metric learning from structured objects (such as strings or trees), because it often involves complex procedures. Most of the work has focused on optimizing a notion of edit distance, which measures (in terms of number of operations) the cost of turning an object into another.We identify two important limitations of current supervised metric learning approaches. First, they allow to improve the performance of local algorithms such as k-nearest neighbors, but metric learning for global algorithms (such as linear classifiers) has not really been studied so far. Second, and perhaps more importantly, the question of the generalization ability of metric learning methods has been largely ignored.In this thesis, we propose theoretical and algorithmic contributions that address these limitations. Our first contribution is the derivation of a new kernel function built from learned edit probabilities. Unlike other string kernels, it is guaranteed to be valid and parameter-free. Our second contribution is a novel framework for learning string and tree edit similarities inspired by the recent theory of (epsilon,gamma,tau)-good similarity functions and formulated as a convex optimization problem. Using uniform stability arguments, we establish theoretical guarantees for the learned similarity that give a bound on the generalization error of a linear classifier built from that similarity. In our third contribution, we extend the same ideas to metric learning from feature vectors by proposing a bilinear similarity learning method that efficiently optimizes the (epsilon,gamma,tau)-goodness. The similarity is learned based on global constraints that are more appropriate to linear classification. Generalization guarantees are derived for our approach, highlighting that our method minimizes a tighter bound on the generalization error of the classifier. Our last contribution is a framework for establishing generalization bounds for a large class of existing metric learning algorithms. It is based on a simple adaptation of the notion of algorithmic robustness and allows the derivation of bounds for various loss functions and regularizers.

La première partie de cette thèse introduit de nouveaux algorithmes de décomposition parcimonieuse de signaux. Basés sur Matching Pursuit (MP) ils répondent au problème suivant : comment réduire le temps de calcul de l'étape de sélection de MP, souvent très coûteuse. En réponse, nous sous-échantillonnons le dictionnaire à chaque itération, en lignes et en colonnes. Nous montrons que cette approche fondée théoriquement affiche de bons résultats en pratique. Nous proposons ensuite un algorithme itératif de descente de gradient par blocs de coordonnées pour sélectionner des caractéristiques en classification multi-classes. Celui-ci s'appuie sur l'utilisation de codes correcteurs d'erreurs transformant le problème en un problème de représentation parcimonieuse simultanée de signaux. La deuxième partie expose de nouvelles inégalités de concentration empiriques de type Bernstein. En premier, elles concernent la théorie des U-statistiques et sont utilisées pour élaborer des bornes en généralisation dans le cadre d'algorithmes de ranking. Ces bornes tirent parti d'un estimateur de variance pour lequel nous proposons un algorithme de calcul efficace. Ensuite, nous présentons une version empirique de l'inégalité de type Bernstein proposée par Freedman [1975] pour les martingales. Ici encore, la force de notre borne réside dans l'introduction d'un estimateur de variance calculable à partir des données. Cela nous permet de proposer des bornes en généralisation pour l'ensemble des algorithmes d'apprentissage en ligne améliorant l'état de l'art et ouvrant la porte à une nouvelle famille d'algorithmes d'apprentissage tirant parti de cette information empirique. / The first part of this thesis introduces new algorithms for the sparse encoding of signals. Based on Matching Pursuit (MP) they focus on the following problem : how to reduce the computation time of the selection step of MP. As an answer, we sub-sample the dictionary in line and column at each iteration. We show that this theoretically grounded approach has good empirical performances. We then propose a bloc coordinate gradient descent algorithm for feature selection problems in the multiclass classification setting. Thanks to the use of error-correcting output codes, this task can be seen as a simultaneous sparse encoding of signals problem. The second part exposes new empirical Bernstein inequalities. Firstly, they concern the theory of the U-Statistics and are applied in order to design generalization bounds for ranking algorithms. These bounds take advantage of a variance estimator and we propose an efficient algorithm to compute it. Then, we present an empirical version of the Bernstein type inequality for martingales by Freedman [1975]. Again, the strength of our result lies in the variance estimator computable from the data. This allows us to propose generalization bounds for online learning algorithms which improve the state of the art and pave the way to a new family of learning algorithms taking advantage of this empirical information.