Dissertations / Theses on the topic 'Apprentissage non-paramétrique'
Create a spot-on reference in APA, MLA, Chicago, Harvard, and other styles
Consult the top 36 dissertations / theses for your research on the topic 'Apprentissage non-paramétrique.'
Next to every source in the list of references, there is an 'Add to bibliography' button. Press on it, and we will generate automatically the bibliographic reference to the chosen work in the citation style you need: APA, MLA, Harvard, Chicago, Vancouver, etc.
You can also download the full text of the academic publication as pdf and read online its abstract whenever available in the metadata.
Browse dissertations / theses on a wide variety of disciplines and organise your bibliography correctly.
Knefati, Muhammad Anas. "Estimation non-paramétrique du quantile conditionnel et apprentissage semi-paramétrique : applications en assurance et actuariat." Thesis, Poitiers, 2015. http://www.theses.fr/2015POIT2280/document.
Full textThe thesis consists of two parts: One part is about the estimation of conditional quantiles and the other is about supervised learning. The "conditional quantile estimate" part is organized into 3 chapters. Chapter 1 is devoted to an introduction to the local linear regression and then goes on to present the methods, the most used in the literature to estimate the smoothing parameter. Chapter 2 addresses the nonparametric estimation methods of conditional quantile and then gives numerical experiments on simulated data and real data. Chapter 3 is devoted to a new conditional quantile estimator, we propose. This estimator is based on the use of asymmetrical kernels w.r.t. x. We show, under some hypothesis, that this new estimator is more efficient than the other estimators already used. The "supervised learning" part is, too, with 3 chapters: Chapter 4 provides an introduction to statistical learning, remembering the basic concepts used in this part. Chapter 5 discusses the conventional methods of supervised classification. Chapter 6 is devoted to propose a method of transferring a semiparametric model. The performance of this method is shown by numerical experiments on morphometric data and credit-scoring data
Lahbib, Dhafer. "Préparation non paramétrique des données pour la fouille de données multi-tables." Phd thesis, Université de Cergy Pontoise, 2012. http://tel.archives-ouvertes.fr/tel-00854142.
Full textSolnon, Matthieu. "Apprentissage statistique multi-tâches." Phd thesis, Université Pierre et Marie Curie - Paris VI, 2013. http://tel.archives-ouvertes.fr/tel-00911498.
Full textScornet, Erwan. "Apprentissage et forêts aléatoires." Thesis, Paris 6, 2015. http://www.theses.fr/2015PA066533/document.
Full textThis is devoted to a nonparametric estimation method called random forests, introduced by Breiman in 2001. Extensively used in a variety of areas, random forests exhibit good empirical performance and can handle massive data sets. However, the mathematical forces driving the algorithm remain largely unknown. After reviewing theoretical literature, we focus on the link between infinite forests (theoretically analyzed) and finite forests (used in practice) aiming at narrowing the gap between theory and practice. In particular, we propose a way to select the number of trees such that the errors of finite and infinite forests are similar. On the other hand, we study quantile forests, a type of algorithms close in spirit to Breiman's forests. In this context, we prove the benefit of trees aggregation: while each tree of quantile forest is not consistent, with a proper subsampling step, the forest is. Next, we show the connection between forests and some particular kernel estimates, which can be made explicit in some cases. We also establish upper bounds on the rate of convergence for these kernel estimates. Then we demonstrate two theorems on the consistency of both pruned and unpruned Breiman forests. We stress the importance of subsampling to demonstrate the consistency of the unpruned Breiman's forests. At last, we present the results of a Dreamchallenge whose goal was to predict the toxicity of several compounds for several patients based on their genetic profile
Lasserre, Marvin. "Apprentissages dans les réseaux bayésiens à base de copules non-paramétriques." Electronic Thesis or Diss., Sorbonne université, 2022. http://www.theses.fr/2022SORUS029.
Full textModeling multivariate continuous distributions is a task of central interest in statistics and machine learning with many applications in science and engineering. However, high-dimensional distributions are difficult to handle and can lead to intractable computations. The Copula Bayesian Networks (CBNs) take advantage of both Bayesian networks (BNs) and copula theory to compactly represent such multivariate distributions. Bayesian networks rely on conditional independences in order to reduce the complexity of the problem, while copula functions allow to model the dependence relation between random variables. The goal of this thesis is to give a common framework to both domains and to propose new learning algorithms for copula Bayesian networks. To do so, we use the fact that CBNs have the same graphical language as BNs which allows us to adapt their learning methods to this model. Moreover, using the empirical Bernstein copula both to design conditional independence tests and to estimate copulas from data, we avoid making parametric assumptions, which gives greater generality to our methods
Genuer, Robin. "Forêts aléatoires : aspects théoriques, sélection de variables et applications." Phd thesis, Université Paris Sud - Paris XI, 2010. http://tel.archives-ouvertes.fr/tel-00550989.
Full textPatra, Benoît. "Apprentissage à "grande échelle" : contribution à l'étude d'algorithmes de clustering répartis asynchrones." Paris 6, 2012. http://www.theses.fr/2012PA066040.
Full textThe subjects addressed in this thesis manuscript are inspired from research problems encountered by the company Lokad, which are summarized in the first chapter. Chapter 2 deals with a nonparametric method for forecasting the quantiles of a real-valued time series. In particular, we establish a consistency result for this technique under minimal assumptions. The remainder of the dissertation is devoted to the analysis of distributed asynchronous clustering algorithms (DALVQ). Chapter 3 first proposes a mathematical description of the models and then offers a theoretical analysis, where the existence of an asymptotical consensus and the almost sure convergence towards critical points of the distortion are proved. In the next chapter, we propose a thorough discussion as well as some experiments on parallelization schemes to be implemented for a practical deployment of DALVQ algorithms. Finally, Chapter 5 contains an effective implementation of DALVQ on the Cloud Computing platform Microsoft Windows Azure. We study, among other topics, the speed ups brought by the algorithm with more parallel computing ressources, and we compare this algorithm with the so-called Lloyd's method, which is also distributed and deployed on Windows Azure
Dallaire, Patrick. "Bayesian nonparametric latent variable models." Doctoral thesis, Université Laval, 2016. http://hdl.handle.net/20.500.11794/26848.
Full textOne of the important problems in machine learning is determining the complexity of the model to learn. Too much complexity leads to overfitting, which finds structures that do not actually exist in the data, while too low complexity leads to underfitting, which means that the expressiveness of the model is insufficient to capture all the structures present in the data. For some probabilistic models, the complexity depends on the introduction of one or more latent variables whose role is to explain the generative process of the data. There are various approaches to identify the appropriate number of latent variables of a model. This thesis covers various Bayesian nonparametric methods capable of determining the number of latent variables to be used and their dimensionality. The popularization of Bayesian nonparametric statistics in the machine learning community is fairly recent. Their main attraction is the fact that they offer highly flexible models and their complexity scales appropriately with the amount of available data. In recent years, research on Bayesian nonparametric learning methods have focused on three main aspects: the construction of new models, the development of inference algorithms and new applications. This thesis presents our contributions to these three topics of research in the context of learning latent variables models. Firstly, we introduce the Pitman-Yor process mixture of Gaussians, a model for learning infinite mixtures of Gaussians. We also present an inference algorithm to discover the latent components of the model and we evaluate it on two practical robotics applications. Our results demonstrate that the proposed approach outperforms, both in performance and flexibility, the traditional learning approaches. Secondly, we propose the extended cascading Indian buffet process, a Bayesian nonparametric probability distribution on the space of directed acyclic graphs. In the context of Bayesian networks, this prior is used to identify the presence of latent variables and the network structure among them. A Markov Chain Monte Carlo inference algorithm is presented and evaluated on structure identification problems and as well as density estimation problems. Lastly, we propose the Indian chefs process, a model more general than the extended cascading Indian buffet process for learning graphs and orders. The advantage of the new model is that it accepts connections among observable variables and it takes into account the order of the variables. We also present a reversible jump Markov Chain Monte Carlo inference algorithm which jointly learns graphs and orders. Experiments are conducted on density estimation problems and testing independence hypotheses. This model is the first Bayesian nonparametric model capable of learning Bayesian learning networks with completely arbitrary graph structures.
Arlot, Sylvain. "Rééchantillonnage et Sélection de modèles." Phd thesis, Université Paris Sud - Paris XI, 2007. http://tel.archives-ouvertes.fr/tel-00198803.
Full textLa majeure partie de ce travail de thèse consiste dans la calibration précise de méthodes de sélection de modèles optimales en pratique, pour le problème de la prédiction. Nous étudions la validation croisée V-fold (très couramment utilisée, mais mal comprise en théorie, notamment pour ce qui est de choisir V) et plusieurs méthodes de pénalisation. Nous proposons des méthodes de calibration précise de pénalités, aussi bien pour ce qui est de leur forme générale que des constantes multiplicatives. L'utilisation du rééchantillonnage permet de résoudre des problèmes difficiles, notamment celui de la régression avec un niveau de bruit variable. Nous validons théoriquement ces méthodes du point de vue non-asymptotique, en prouvant des inégalités oracle et des propriétés d'adaptation. Ces résultats reposent entre autres sur des inégalités de concentration.
Un second problème que nous abordons est celui des régions de confiance et des tests multiples, lorsque l'on dispose d'observations de grande dimension, présentant des corrélations générales et inconnues. L'utilisation de méthodes de rééchantillonnage permet de s'affranchir du fléau de la dimension, et d'"apprendre" ces corrélations. Nous proposons principalement deux méthodes, et prouvons pour chacune un contrôle non-asymptotique de leur niveau.
Averyanov, Yaroslav. "Concevoir et analyser de nouvelles règles d’arrêt prématuré pour économiser les ressources de calcul." Thesis, Lille 1, 2020. http://www.theses.fr/2020LIL1I048.
Full textThis work develops and analyzes strategies for constructing instances of the so-called early stopping rules applied to some iterative learning algorithms for estimating the regression function. Such quantities are data-driven rules indicating when to stop the iterative learning process to reach a trade-off between computational costs and the statistical precision. Unlike a large part of the existing literature on early stopping, where these rules only depend on the data in a "weak manner", we provide data-driven solutions for the aforementioned problem without utilizing validation data. The crucial idea exploited here is that of the minimum discrepancy principle (MDP), which shows when to stop an iterative learning algorithm. To the best of our knowledge, this idea dates back to the work of Vladimir A. Morozov in the 1960s-1970s who studied linear ill-posed problems and their regularization, mostly inspired by mathematical physics problems. Among different applications of this line of work, the so-called spectral filter estimators such as spectral cut-off, Landweber iterations, and Tikhonov (ridge) regularization have received quite a lot of attention (e.g., in statistical inverse problems). It is worth mentioning that the minimum discrepancy principle consists in controlling the residuals of an estimator (which are iteratively minimized) and properly setting a threshold for them such that one can achieve some (minimax) optimality. The first part of this thesis is dedicated to theoretical guarantees of stopping rules based on the minimum discrepancy principle and applied to gradient descent, and Tikhonov (ridge) regression in the framework of reproducing kernel Hilbert space (RKHS). There, we show that this principle provides a minimax optimal functional estimator of the regression function when the rank of the kernel is finite. However, when one deals with infinite-rank reproducing kernels, the resulting estimator will be only suboptimal. While looking for a solution, we found the existence of the so-called residuals polynomial smoothing strategy. This strategy (combined with MDP) has been proved to be optimal for the spectral cut-off estimator in the linear Gaussian sequence model. We borrow this strategy, modify the stopping rule accordingly, and prove that the smoothed minimum discrepancy principle yields a minimax optimal functional estimator over a range of function spaces, which includes the well-known Sobolev function class. Our second contribution consists in exploring the theoretical properties of the minimum discrepancy stopping rule applied to the more general family of linear estimators. The main difficulty of this approach is that, unlike the spectral filter estimators considered earlier, linear estimators do no longer lead to monotonic quantities (the bias and variance terms). Let us mention that this is also the case for famous algorithms such as Stochastic Gradient Descent. Motivated by further practical applications, we work with the widely used k-NN regression estimator as a reliable first example. We prove that the aforementioned stopping rule leads to a minimax optimal functional estimator, in particular, over the class of Lipschitz functions on a bounded domain.The third contribution consists in illustrating through empirical experiments that for choosing the tuning parameter in a linear estimator (the k-NN regression, Nadaraya-Watson, and variable selection estimators), the MDP-based early stopping rule performs comparably well with respect to other widely used and known model selection criteria
Löser, Kevin. "Apprentissage non-supervisé de la morphologie des langues à l’aide de modèles bayésiens non-paramétriques." Thesis, Université Paris-Saclay (ComUE), 2019. http://www.theses.fr/2019SACLS203/document.
Full textA crucial issue in statistical natural language processing is the issue of sparsity, namely the fact that in a given learning corpus, most linguistic events have low occurrence frequencies, and that an infinite number of structures allowed by a language will not be observed in the corpus. Neural models have already contributed to solving this issue by inferring continuous word representations. These continuous representations allow to structure the lexicon by inducing semantic or syntactic similarity between words. However, current neural models only partially solve the sparsity issue, due to the fact that they require a vectorial representation for every word in the lexicon, but are unable to infer sensible representations for unseen words. This issue is especially present in morphologically rich languages, where word formation processes yield a proliferation of possible word forms, and little overlap between the lexicon observed during model training, and the lexicon encountered during its use. Today, several languages are used on the Web besides English, and engineering translation systems that can handle morphologies that are very different from western European languages has become a major stake. The goal of this thesis is to develop new statistical models that are able to infer in an unsupervised fashion the word formation processes underlying an observed lexicon, in order to produce morphological analyses of new unseen word forms
Dang, Hong-Phuong. "Approches bayésiennes non paramétriques et apprentissage de dictionnaire pour les problèmes inverses en traitement d'image." Thesis, Ecole centrale de Lille, 2016. http://www.theses.fr/2016ECLI0019/document.
Full textDictionary learning for sparse representation has been widely advocated for solving inverse problems. Optimization methods and parametric approaches towards dictionary learning have been particularly explored. These methods meet some limitations, particularly related to the choice of parameters. In general, the dictionary size is fixed in advance, and sparsity or noise level may also be needed. In this thesis, we show how to perform jointly dictionary and parameter learning, with an emphasis on image processing. We propose and study the Indian Buffet Process for Dictionary Learning (IBP-DL) method, using a bayesian nonparametric approach.A primer on bayesian nonparametrics is first presented. Dirichlet and Beta processes and their respective derivatives, the Chinese restaurant and Indian Buffet processes are described. The proposed model for dictionary learning relies on an Indian Buffet prior, which permits to learn an adaptive size dictionary. The Monte-Carlo method for inference is detailed. Noise and sparsity levels are also inferred, so that in practice no parameter tuning is required. Numerical experiments illustrate the performances of the approach in different settings: image denoising, inpainting and compressed sensing. Results are compared with state-of-the art methods is made. Matlab and C sources are available for sake of reproducibility
Avalos, Marta. "Modèles additifs parcimonieux." Phd thesis, Université de Technologie de Compiègne, 2004. http://tel.archives-ouvertes.fr/tel-00008802.
Full textBartcus, Marius. "Bayesian non-parametric parsimonious mixtures for model-based clustering." Thesis, Toulon, 2015. http://www.theses.fr/2015TOUL0010/document.
Full textThis thesis focuses on statistical learning and multi-dimensional data analysis. It particularly focuses on unsupervised learning of generative models for model-based clustering. We study the Gaussians mixture models, in the context of maximum likelihood estimation via the EM algorithm, as well as in the Bayesian estimation context by maximum a posteriori via Markov Chain Monte Carlo (MCMC) sampling techniques. We mainly consider the parsimonious mixture models which are based on a spectral decomposition of the covariance matrix and provide a flexible framework particularly for the analysis of high-dimensional data. Then, we investigate non-parametric Bayesian mixtures which are based on general flexible processes such as the Dirichlet process and the Chinese Restaurant Process. This non-parametric model formulation is relevant for both learning the model, as well for dealing with the issue of model selection. We propose new Bayesian non-parametric parsimonious mixtures and derive a MCMC sampling technique where the mixture model and the number of mixture components are simultaneously learned from the data. The selection of the model structure is performed by using Bayes Factors. These models, by their non-parametric and sparse formulation, are useful for the analysis of large data sets when the number of classes is undetermined and increases with the data, and when the dimension is high. The models are validated on simulated data and standard real data sets. Then, they are applied to a real difficult problem of automatic structuring of complex bioacoustic data issued from whale song signals. Finally, we open Markovian perspectives via hierarchical Dirichlet processes hidden Markov models
Sibony, Eric. "Analyse mustirésolution de données de classements." Electronic Thesis or Diss., Paris, ENST, 2016. http://www.theses.fr/2016ENST0036.
Full textThis thesis introduces a multiresolution analysis framework for ranking data. Initiated in the 18th century in the context of elections, the analysis of ranking data has attracted a major interest in many fields of the scientific literature : psychometry, statistics, economics, operations research, machine learning or computational social choice among others. It has been even more revitalized by modern applications such as recommender systems, where the goal is to infer users preferences in order to make them the best personalized suggestions. In these settings, users express their preferences only on small and varying subsets of a large catalog of items. The analysis of such incomplete rankings poses however both a great statistical and computational challenge, leading industrial actors to use methods that only exploit a fraction of available information. This thesis introduces a new representation for the data, which by construction overcomes the two aforementioned challenges. Though it relies on results from combinatorics and algebraic topology, it shares several analogies with multiresolution analysis, offering a natural and efficient framework for the analysis of incomplete rankings. As it does not involve any assumption on the data, it already leads to overperforming estimators in small-scale settings and can be combined with many regularization procedures for large-scale settings. For all those reasons, we believe that this multiresolution representation paves the way for a wide range of future developments and applications
Mahler, Nicolas. "Machine learning methods for discrete multi-scale fows : application to finance." Phd thesis, École normale supérieure de Cachan - ENS Cachan, 2012. http://tel.archives-ouvertes.fr/tel-00749717.
Full textRouvière, Laurent. "Estimation de densité en dimension élevée et classification de courbes." Phd thesis, Université Montpellier II - Sciences et Techniques du Languedoc, 2005. http://tel.archives-ouvertes.fr/tel-00011624.
Full textLa première partie, intitulée compléments sur les histogrammes modifiés, est composée de deux chapitres consacrés l'étude d'une famille d'estimateurs non paramétriques de la densité, les histogrammes modifiés, connus pour posséder de bonnes propriétés de convergence au sens des critères de la théorie de l'information. Dans le premier chapitre, ces estimateurs sont envisagés comme des systèmes dynamiques espace d'états de dimension infinie. Le second chapitre est consacré l'étude de ces estimateurs pour des dimensions suprieures un.
La deuxième partie de la thèse, intituleé méthodes combinatoires en estimation de la densité, se divise en deux chapitres. Nous nous intéressons dans cette partie aux performances distance finie d'estimateurs de la densité sélectionnés à l'intérieur d'une famille d'estimateurs candidats, dont le cardinal n'est pas nécessairement fini. Dans le premier chapitre, nous étudions les performances de ces méthodes dans le cadre de la sélection des différents paramètres des histogrammes modifiés. Nous poursuivons, dans le deuxième chapitre, par la sélection d'estimateurs à noyau dont le paramètre de lissage s'adapte localement au point d'estimation et aux données.
Enfin, la troisième et dernière partie, plus appliquée et indépendante des précédentes, présente une nouvelle méthode permettant de classer des courbes partir d'une décomposition des observations dans des bases d'ondelettes.
Wira, Patrice. "Approches neuromimétiques pour l'identification et la commande." Habilitation à diriger des recherches, Université de Haute Alsace - Mulhouse, 2009. http://tel.archives-ouvertes.fr/tel-00605218.
Full textTran, Gia-Lac. "Advances in Deep Gaussian Processes : calibration and sparsification." Electronic Thesis or Diss., Sorbonne université, 2020. https://accesdistant.sorbonne-universite.fr/login?url=https://theses-intra.sorbonne-universite.fr/2020SORUS410.pdf.
Full textGaussian Processes (GPs) are an attractive specific way of doing non-parametric Bayesian modeling in a supervised learning problem. It is well-known that GPs are able to make inferences as well as predictive uncertainties with a firm mathematical background. However, GPs are often unfavorable by the practitioners due to their kernel's expressiveness and the computational requirements. Integration of (convolutional) neural networks and GPs are a promising solution to enhance the representational power. As our first contribution, we empirically show that these combinations are miscalibrated, which leads to over-confident predictions. We also propose a novel well-calibrated solution to merge neural structures and GPs by using random features and variational inference techniques. In addition, these frameworks can be intuitively extended to reduce the computational cost by using structural random features. In terms of computational cost, the exact Gaussian Processes require the cubic complexity to training size. Inducing point-based Gaussian Processes are a common choice to mitigate the bottleneck by selecting a small set of active points through a global distillation from available observations. However, the general case remains elusive and it is still possible that the required number of active points may exceed a certain computational budget. In our second study, we propose Sparse-within-Sparse Gaussian Processes which enable the approximation with a large number of inducing points without suffering a prohibitive computational cost
Khaleghi, Azadeh. "Sur quelques problèmes non-supervisés impliquant des séries temporelles hautement dépendantes." Phd thesis, Université des Sciences et Technologie de Lille - Lille I, 2013. http://tel.archives-ouvertes.fr/tel-00920333.
Full textBaelde, Maxime. "Modèles génératifs pour la classification et la séparation de sources sonores en temps-réel." Thesis, Lille 1, 2019. http://www.theses.fr/2019LIL1I058/document.
Full textThis thesis is part of the A-Volute company, an audio enhancement softwares editor. It offers a radar that translates multi-channel audio information into visual information in real-time. This radar, although relevant, lacks intelligence because it only analyses the audio stream in terms of energy and not in terms of separate sound sources. The purpose of this thesis is to develop algorithms for classifying and separating sound sources in real time. On the one hand, audio source classification aims to assign a label (e.g. voice) to a monophonic (one label) or polyphonic (several labels) sound. The developed method uses a specific feature, the normalized power spectrum, which is useful in both monophonic and polyphonic cases due to its additive properties of the sound sources. This method uses a generative model that allows to derive a decision rule based on a non-parametric estimation. The real-time constraint is achieved by pre-processing the prototypes with a hierarchical clustering. The results are encouraging on different databases (owned and benchmark), both in terms of accuracy and computation time, especially in the polyphonic case. On the other hand, source separation consists in estimating the sources in terms of signal in a mixture. Two approaches to this purpose were considered in this thesis. The first considers the signals to be found as missing data and estimates them through a generative process and probabilistic modelling. The other approach consists, from sound examples present in a database, in computing optimal transformations of several examples whose combination tends towards the observed mixture. The two proposals are complementary, each having advantages and drawbacks (computation time for the first, interpretability of the result for the second). The experimental results seem promising and allow us to consider interesting research perspectives for each of the proposals
Sibony, Eric. "Analyse mustirésolution de données de classements." Thesis, Paris, ENST, 2016. http://www.theses.fr/2016ENST0036/document.
Full textThis thesis introduces a multiresolution analysis framework for ranking data. Initiated in the 18th century in the context of elections, the analysis of ranking data has attracted a major interest in many fields of the scientific literature : psychometry, statistics, economics, operations research, machine learning or computational social choice among others. It has been even more revitalized by modern applications such as recommender systems, where the goal is to infer users preferences in order to make them the best personalized suggestions. In these settings, users express their preferences only on small and varying subsets of a large catalog of items. The analysis of such incomplete rankings poses however both a great statistical and computational challenge, leading industrial actors to use methods that only exploit a fraction of available information. This thesis introduces a new representation for the data, which by construction overcomes the two aforementioned challenges. Though it relies on results from combinatorics and algebraic topology, it shares several analogies with multiresolution analysis, offering a natural and efficient framework for the analysis of incomplete rankings. As it does not involve any assumption on the data, it already leads to overperforming estimators in small-scale settings and can be combined with many regularization procedures for large-scale settings. For all those reasons, we believe that this multiresolution representation paves the way for a wide range of future developments and applications
Calandriello, Daniele. "Efficient sequential learning in structured and constrained environments." Thesis, Lille 1, 2017. http://www.theses.fr/2017LIL10216/document.
Full textThe main advantage of non-parametric models is that the accuracy of the model (degrees of freedom) adapts to the number of samples. The main drawback is the so-called "curse of kernelization": to learn the model we must first compute a similarity matrix among all samples, which requires quadratic space and time and is unfeasible for large datasets. Nonetheless the underlying effective dimension (effective d.o.f.) of the dataset is often much smaller than its size, and we can replace the dataset with a subset (dictionary) of highly informative samples. Unfortunately, fast data-oblivious selection methods (e.g., uniform sampling) almost always discard useful information, while data-adaptive methods that provably construct an accurate dictionary, such as ridge leverage score (RLS) sampling, have a quadratic time/space cost. In this thesis we introduce a new single-pass streaming RLS sampling approach that sequentially construct the dictionary, where each step compares a new sample only with the current intermediate dictionary and not all past samples. We prove that the size of all intermediate dictionaries scales only with the effective dimension of the dataset, and therefore guarantee a per-step time and space complexity independent from the number of samples. This reduces the overall time required to construct provably accurate dictionaries from quadratic to near-linear, or even logarithmic when parallelized. Finally, for many non-parametric learning problems (e.g., K-PCA, graph SSL, online kernel learning) we we show that we can can use the generated dictionaries to compute approximate solutions in near-linear that are both provably accurate and empirically competitive
Dieuleveut, Aymeric. "Stochastic approximation in Hilbert spaces." Thesis, Paris Sciences et Lettres (ComUE), 2017. http://www.theses.fr/2017PSLEE059/document.
Full textThe goal of supervised machine learning is to infer relationships between a phenomenon one seeks to predict and “explanatory” variables. To that end, multiple occurrences of the phenomenon are observed, from which a prediction rule is constructed. The last two decades have witnessed the apparition of very large data-sets, both in terms of the number of observations (e.g., in image analysis) and in terms of the number of explanatory variables (e.g., in genetics). This has raised two challenges: first, avoiding the pitfall of over-fitting, especially when the number of explanatory variables is much higher than the number of observations; and second, dealing with the computational constraints, such as when the mere resolution of a linear system becomes a difficulty of its own. Algorithms that take their roots in stochastic approximation methods tackle both of these difficulties simultaneously: these stochastic methods dramatically reduce the computational cost, without degrading the quality of the proposed prediction rule, and they can naturally avoid over-fitting. As a consequence, the core of this thesis will be the study of stochastic gradient methods. The popular parametric methods give predictors which are linear functions of a set ofexplanatory variables. However, they often result in an imprecise approximation of the underlying statistical structure. In the non-parametric setting, which is paramount in this thesis, this restriction is lifted. The class of functions from which the predictor is proposed depends on the observations. In practice, these methods have multiple purposes, and are essential for learning with non-vectorial data, which can be mapped onto a vector in a functional space using a positive definite kernel. This allows to use algorithms designed for vectorial data, but requires the analysis to be made in the non-parametric associated space: the reproducing kernel Hilbert space. Moreover, the analysis of non-parametric regression also sheds some light on the parametric setting when the number of predictors is much larger than the number of observations. The first contribution of this thesis is to provide a detailed analysis of stochastic approximation in the non-parametric setting, precisely in reproducing kernel Hilbert spaces. This analysis proves optimal convergence rates for the averaged stochastic gradient descent algorithm. As we take special care in using minimal assumptions, it applies to numerous situations, and covers both the settings in which the number of observations is known a priori, and situations in which the learning algorithm works in an on-line fashion. The second contribution is an algorithm based on acceleration, which converges at optimal speed, both from the optimization point of view and from the statistical one. In the non-parametric setting, this can improve the convergence rate up to optimality, even inparticular regimes for which the first algorithm remains sub-optimal. Finally, the third contribution of the thesis consists in an extension of the framework beyond the least-square loss. The stochastic gradient descent algorithm is analyzed as a Markov chain. This point of view leads to an intuitive and insightful interpretation, that outlines the differences between the quadratic setting and the more general setting. A simple method resulting in provable improvements in the convergence is then proposed
Carriere, Mathieu. "On Metric and Statistical Properties of Topological Descriptors for geometric Data." Thesis, Université Paris-Saclay (ComUE), 2017. http://www.theses.fr/2017SACLS433/document.
Full textIn the context of supervised Machine Learning, finding alternate representations, or descriptors, for data is of primary interest since it can greatly enhance the performance of algorithms. Among them, topological descriptors focus on and encode the topological information contained in geometric data. One advantage of using these descriptors is that they enjoy many good and desireable properties, due to their topological nature. For instance, they are invariant to continuous deformations of data. However, the main drawback of these descriptors is that they often lack the structure and operations required by most Machine Learning algorithms, such as a means or scalar products. In this thesis, we study the metric and statistical properties of the most common topological descriptors, the persistence diagrams and the Mappers. In particular, we show that the Mapper, which is empirically instable, can be stabilized with an appropriate metric, that we use later on to conpute confidence regions and automatic tuning of its parameters. Concerning persistence diagrams, we show that scalar products can be defined with kernel methods by defining two kernels, or embeddings, into finite and infinite dimensional Hilbert spaces
Godard, Pierre. "Unsupervised word discovery for computational language documentation." Thesis, Université Paris-Saclay (ComUE), 2019. http://www.theses.fr/2019SACLS062/document.
Full textLanguage diversity is under considerable pressure: half of the world’s languages could disappear by the end of this century. This realization has sparked many initiatives in documentary linguistics in the past two decades, and 2019 has been proclaimed the International Year of Indigenous Languages by the United Nations, to raise public awareness of the issue and foster initiatives for language documentation and preservation. Yet documentation and preservation are time-consuming processes, and the supply of field linguists is limited. Consequently, the emerging field of computational language documentation (CLD) seeks to assist linguists in providing them with automatic processing tools. The Breaking the Unwritten Language Barrier (BULB) project, for instance, constitutes one of the efforts defining this new field, bringing together linguists and computer scientists. This thesis examines the particular problem of discovering words in an unsegmented stream of characters, or phonemes, transcribed from speech in a very-low-resource setting. This primarily involves a segmentation procedure, which can also be paired with an alignment procedure when a translation is available. Using two realistic Bantu corpora for language documentation, one in Mboshi (Republic of the Congo) and the other in Myene (Gabon), we benchmark various monolingual and bilingual unsupervised word discovery methods. We then show that using expert knowledge in the Adaptor Grammar framework can vastly improve segmentation results, and we indicate ways to use this framework as a decision tool for the linguist. We also propose a tonal variant for a strong nonparametric Bayesian segmentation algorithm, making use of a modified backoff scheme designed to capture tonal structure. To leverage the weak supervision given by a translation, we finally propose and extend an attention-based neural segmentation method, improving significantly the segmentation performance of an existing bilingual method
Dulac, Adrien. "Etude des modèles à composition mixée pour l'analyse de réseaux complexes." Thesis, Université Grenoble Alpes (ComUE), 2018. http://www.theses.fr/2018GREAM080/document.
Full textRelational data are ubiquitous in the nature and their accessibility has not ceased to increase in recent years. Those data, see as a whole, form a network, which can be represented by a data structure called a graph, where each vertex of the graph is an entity and each edge a connection between pair of vertices. Complex networks in general, such as the Web, communication networks or social network, are known to exhibit common structural properties that emerge through their graphs. In this work we emphasize two important properties called *homophilly* and *preferential attachment* that arise on most of the real-world networks. We firstly study a class of powerful *random graph models* in a Bayesian nonparametric setting, called *mixed-membership model* and we focus on showing whether the models in this class comply with the mentioned properties, after giving formal definitions in a probabilistic context of the latter. Furthermore, we empirically evaluate our findings on synthetic and real-world network datasets. Secondly, we propose a new model, which extends the former Stochastic Mixed-Membership Model, for weighted networks and we develop an efficient inference algorithm able to scale to large-scale networks
Prendes, Jorge. "New statistical modeling of multi-sensor images with application to change detection." Thesis, Université Paris-Saclay (ComUE), 2015. http://www.theses.fr/2015SACLC006/document.
Full textRemote sensing images are images of the Earth surface acquired from satellites or air-borne equipment. These images are becoming widely available nowadays and its sensor technology is evolving fast. Classical sensors are improving in terms of resolution and noise level, while new kinds of sensors are proving to be useful. Multispectral image sensors are standard nowadays and synthetic aperture radar (SAR) images are very popular.The availability of different kind of sensors is very advantageous since it allows us to capture a wide variety of properties of the objects contained in a scene. These properties can be exploited to extract richer information about these objects. One of the main applications of remote sensing images is the detection of changes in multitemporal datasets (images of the same area acquired at different times). Change detection for images acquired by homogeneous sensors has been of interest for a long time. However the wide range of different sensors found in remote sensing makes the detection of changes in images acquired by heterogeneous sensors an interesting challenge.Accurate change detectors adapted to heterogeneous sensors are needed for the management of natural disasters. Databases of optical images are readily available for an extensive catalog of locations, but, good climate conditions and daylight are required to capture them. On the other hand, SAR images can be quickly captured, regardless of the weather conditions or the daytime. For these reasons, optical and SAR images are of specific interest for tracking natural disasters, by detecting the changes before and after the event.The main interest of this thesis is to study statistical approaches to detect changes in images acquired by heterogeneous sensors. Chapter 1 presents an introduction to remote sensing images. It also briefly reviews the different change detection methods proposed in the literature. Additionally, this chapter presents the motivation to detect changes between heterogeneous sensors and its difficulties.Chapter 2 studies the statistical properties of co-registered images in the absence of change, in particular for optical and SAR images. In this chapter a finite mixture model is proposed to describe the statistics of these images. The performance of classical statistical change detection methods is also studied by taking into account the proposed statistical model. In several situations it is found that these classical methods fail for change detection.Chapter 3 studies the properties of the parameters associated with the proposed statistical mixture model. We assume that the model parameters belong to a manifold in the absence of change, which is then used to construct a new similarity measure overcoming the limitations of classic statistical approaches. Furthermore, an approach to estimate the proposed similarity measure is described. Finally, the proposed change detection strategy is validated on synthetic images and compared with previous strategies.Chapter 4 studies Bayesian non parametric algorithm to improve the estimation of the proposed similarity measure. This algorithm is based on a Chinese restaurant process and a Markov random field taking advantage of the spatial correlations between adjacent pixels of the image. This chapter also defines a new Jeffreys prior for the concentration parameter of this Chinese restaurant process. The estimation of the different model parameters is conducted using a collapsed Gibbs sampler. The proposed strategy is validated on synthetic images and compared with the previously proposed strategy. Finally, Chapter 5 is dedicated to the validation of the proposed change detection framework on real datasets, where encouraging results are obtained in all cases. Including the Bayesian non parametric model into the change detection strategy improves change detection performance at the expenses of an increased computational cost
Kamari, Halaleh. "Qualité prédictive des méta-modèles construits sur des espaces de Hilbert à noyau auto-reproduisant et analyse de sensibilité des modèles complexes." Thesis, université Paris-Saclay, 2020. http://www.theses.fr/2020UPASE010.
Full textIn this work, the problem of estimating a meta-model of a complex model, denoted m, is considered. The model m depends on d input variables X1 , ..., Xd that are independent and have a known law. The meta-model, denoted f ∗ , approximates the Hoeffding decomposition of m, and allows to estimate its Sobol indices. It belongs to a reproducing kernel Hilbert space (RKHS), denoted H, which is constructed as a direct sum of Hilbert spaces (Durrande et al. (2013)). The estimator of the meta-model, denoted f^, is calculated by minimizing a least-squares criterion penalized by the sum of the Hilbert norm and the empirical L2-norm (Huet and Taupin (2017)). This procedure, called RKHS ridge group sparse, allows both to select and estimate the terms in the Hoeffding decomposition, and therefore, to select the Sobol indices that are non-zero and estimate them. It makes possible to estimate the Sobol indices even of high order, a point known to be difficult in practice.This work consists of a theoretical part and a practical part. In the theoretical part, I established upper bounds of the empirical L2 risk and the L2 risk of the estimator f^. That is, upper bounds with respect to the L2-norm and the empirical L2-norm for the f^ distance between the model m and its estimation f into the RKHS H. In the practical part, I developed an R package, called RKHSMetaMod, that implements the RKHS ridge group sparse procedure and a spacial case of it called the RKHS group lasso procedure. This package can be applied to a known model that is calculable in all points or an unknown regression model. In order to optimize the execution time and the storage memory, except for a function that is written in R, all of the functions of the RKHSMetaMod package are written using C++ libraries GSL and Eigen. These functions are then interfaced with the R environment in order to propose an user friendly package. The performance of the package functions in terms of the predictive quality of the estimator and the estimation of the Sobol indices, is validated by a simulation study
Depecker, Marine. "Méthodes d'apprentissage statistique pour le scoring." Phd thesis, Télécom ParisTech, 2010. http://pastel.archives-ouvertes.fr/pastel-00572421.
Full textDepecker, Marine. "Méthodes d'apprentissage statistique pour le scoring." Phd thesis, Paris, Télécom ParisTech, 2010. https://pastel.hal.science/pastel-00572421.
Full textBipartite ranking is a statistical issue consisting in sorting objects lying in a multidimensional feature space, randomly associated with binary labels, so that positive instances appear on top of the list with highest probability. This research work aims at developing a tree-induction ranking method based on a top-down recursive partitioning strategy and leading to a scoring function summarized by a rooted, binary, left-right oriented tree graph. In order to improve the flexibility of this learning method, we introduce a partition-based procedure involving complex and adaptive splitting rules. We then tackle the classical issue of model selection and propose two penalization-based procedures providing the best ranking tree for prediction. Finally, in order to reduce the instability of ranking trees and increase their accuracy, we propose to adapt two re-sampling and aggregating procedures introduced by Breiman in the classification and regression contexts: bagging (1996) and random forests (2001). An empirical comparison between several versions of this ranking algorithm and state-of-the-art scoring methods is provided. We also present the results output on industrial objectivization data. Last but not least, we introduce a two-stage testing procedure aiming at solving the two-sample problem in a multidimensional setting, based on the proposed ranking algorithm and on one-dimensional rank tests
Duchemin, Quentin. "Growth dynamics of large networks using hidden Markov chains." Thesis, Université Gustave Eiffel, 2022. https://tel.archives-ouvertes.fr/tel-03749513.
Full textThe first part of this thesis aims at introducing new models of random graphs that account for the temporal evolution of networks. More precisely, we focus on growth models where at each instant a new node is added to the existing graph. We attribute to this new entrant properties that characterize its connectivity to the rest of the network and these properties depend only on the previously introduced node. Our random graph models are thus governed by a latent Markovian dynamic characterizing the sequence of nodes in the graph. We are particularly interested in the Stochastic Block Model and in Random Geometric Graphs for which we propose algorithms to estimate the unknown parameters or functions defining the model. We then show how these estimates allow us to solve link prediction or collaborative filtering problems in networks.The theoretical analysis of the above-mentioned algorithms requires advanced probabilistic tools. In particular, one of our proof is relying on a concentration inequality for U-statistics in a dependent framework. Few papers have addressed this thorny question and existing works consider sets of assumptions that do not meet our needs. Therefore, the second part of this manuscript will be devoted to the proof of a concentration inequality for U-statistics of order two for uniformly ergodic Markov chains. In Chapter 5, we exploit this concentration result for U-statistics to make new contributions to three very active areas of Statistics and Machine Learning.Still motivated by link prediction problems in graphs, we study post-selection inference procedures in the framework of logistic regression with $L^1$ penalty. We prove a central limit theorem under the distribution conditional on the selection event and derive asymptotically valid testing procedures and confidence intervals
Hebbal, Ali. "Deep gaussian processes for the analysis and optimization of complex systems : application to aerospace system design." Thesis, Lille, 2021. http://www.theses.fr/2021LILUI016.
Full textIn engineering, the design of complex systems, such as aerospace launch vehicles, involves the analysis and optimization of problems presenting diverse challenges. Actually, the designer has to take into account different aspects in the design of complex systems, such as the presence of black-box computationally expensive functions, the complex behavior of the optimized performance (e.g., abrupt change of a physical property here referred as non-stationarity), the multiple objectives and constraints involved, the multi-source information handling in a multi-fidelity framework, and the epistemic and aleatory uncertainties affecting the physical models. A wide range of machine learning methods are used to address these various challenges. Among these approaches, Gaussian Processes (GPs), benefiting from their Bayesian and non-parametric formulation, are popular in the literature and diverse state-of-the-art algorithms for the design of complex systems are based on these models.Despite being widely used for the analysis and optimization of complex systems, GPs, still present some limitations. For the optimization of computationally expensive functions, GPs are used within the Bayesian optimization framework as regression models. However, for the optimization of non-stationary problems, they are not suitable due to the use of a prior stationary covariance function. Furthermore, in Bayesian optimization of multiple objectives, a GP is used for each involved objective independently, which prevents the exhibition of a potential correlation between the objectives. Another limitation occurs in multi-fidelity analysis where GP-based models are used to improve high-fidelity models using low-fidelity information. However, these models usually assume that the different fidelity input spaces are identically defined, which is not the case in some design problems.In this thesis, approaches are developed to overcome the limits of GPs in the analysis and optimization of complex systems. These approaches are based on Deep Gaussian Processes (DGPs), the hierarchical generalization of Gaussian processes.To handle non-stationarity in Bayesian optimization, a framework is developed that couples Bayesian optimization with DGPs. The inner layers allow a non-parametric Bayesian mapping of the input space to better represent non-stationary functions. For multi-objective Bayesian optimization, a multi-objective DGP model is developed. Each layer of this model corresponds to an objective and the different layers are connected with undirected edges to encode the potential correlation between objectives. Moreover, a computational approach for the expected hyper-volume improvement is proposed to take into account this correlation at the infill criterion level as well. Finally, to address multi-fidelity analysis for different input space definitions, a two-level DGP model is developed. This model allows a joint optimization of the multi-fidelity model and the input space mapping between fidelities.The different approaches developed are assessed on analytical problems as well as on representative aerospace vehicle design problems with respect to state-of-the-art approaches
Laouti, Nassim. "Diagnostic de défauts par les Machines à Vecteurs Supports : application à différents systèmes mutivariables nonlinéaires." Phd thesis, Université Claude Bernard - Lyon I, 2012. http://tel.archives-ouvertes.fr/tel-00985437.
Full textLoth, Manuel. "Algorithmes d'Ensemble Actif pour le LASSO." Phd thesis, Université des Sciences et Technologie de Lille - Lille I, 2011. http://tel.archives-ouvertes.fr/tel-00845441.
Full textSenécal, Jean-Sébastien. "Accélérer l'entraînement d'un modèle non-paramétrique de densité non normalisée par échantillonnage aléatoire." Thèse, 2003. http://hdl.handle.net/1866/14518.
Full text