The development of modern information technology has enabled collecting data of unprecedented size and complexity. Examples include web text data, microarray & proteomics, and data from scientific domains (e.g., meteorology). To learn from these high dimensional and complex data, traditional machine learning techniques often suffer from the curse of dimensionality and unaffordable computational cost. However, learning from large-scale high-dimensional data promises big payoffs in text mining, gene analysis, and numerous other consequential tasks. Recently developed sparse learning techniques provide us a suite of tools for understanding and exploring high dimensional data from many areas in science and engineering. By exploring sparsity, we can always learn a parsimonious and compact model which is more interpretable and computationally tractable at application time. When it is known that the underlying model is indeed sparse, sparse learning methods can provide us a more consistent model and much improved prediction performance. However, the existing methods are still insufficient for modeling complex or dynamic structures of the data, such as those evidenced in pathways of genomic data, gene regulatory network, and synonyms in text data. This thesis develops structured sparse learning methods along with scalable optimization algorithms to explore and predict high dimensional data with complex structures. In particular, we address three aspects of structured sparse learning: 1. Efficient and scalable optimization methods with fast convergence guarantees for a wide spectrum of high-dimensional learning tasks, including single or multi-task structured regression, canonical correlation analysis as well as online sparse learning. 2. Learning dynamic structures of different types of undirected graphical models, e.g., conditional Gaussian or conditional forest graphical models. 3. Demonstrating the usefulness of the proposed methods in various applications, e.g., computational genomics and spatial-temporal climatological data. In addition, we also design specialized sparse learning methods for text mining applications, including ranking and latent semantic analysis. In the last part of the thesis, we also present the future direction of the high-dimensional structured sparse learning from both computational and statistical aspects.

The mission of the research presented in this thesis is to give computers the power to sense and react to human activities. Without the ability to sense the surroundings and understand what humans are doing, computers will not be able to provide active, timely, appropriate, and considerate services to the humans. To accomplish this mission, the work stands on the shoulders of two giants: Machine learning and ubiquitous computing. Because of the ubiquity of sensor-enabled mobile and wearable devices, there has been an emerging opportunity to sense, learn, and infer human activities from the sensor data by leveraging state-of-the-art machine learning algorithms. While having shown promising results in human activity recognition, most existing approaches using supervised or semi-supervised learning have two fundamental problems. Firstly, most existing approaches require a large set of labeled sensor data for every target class, which requires a costly effort from human annotators. Secondly, an unseen new activity cannot be recognized if no training samples of that activity are available in the dataset. In light of these problems, a new approach in this area is proposed in our research. This thesis presents our novel approach to address the problem of human activity recognition when few or no training samples of the target activities are available. The main hypothesis is that the problem can be solved by the proposed NuActiv activity recognition framework, which consists of modeling the hierarchical and sequential structure of human activities, as well as bringing humans in the loop of model training. By injecting human knowledge about the hierarchical nature of human activities, a semantic attribute representation and a two-layer attribute-based learning approach are designed. To model the sequential structure, a probabilistic graphical model is further proposed to take into account the temporal dependency of activities and attributes. Finally, an active learning algorithm is developed to reinforce the recognition accuracy using minimal user feedback. The hypothesis and approaches presented in this thesis are validated by two case studies and real-world experiments on exercise activities and daily life activities. Experimental results show that the NuActiv framework can effectively recognize unseen new activities even without any training data, with up to 70-80% precision and recall rate. It also outperforms supervised learning with limited labeled data for the new classes. The results significantly advance the state of the art in human activity recognition, and represent a promising step towards bridging the gap between computers and humans.

Recent experimental techniques in biology made possible the acquisition of overwhelming amounts of data concerning complex biological networks, such as neural networks, gene regulation networks and protein-protein interaction networks. These techniques are able to record states of individual components of such networks (neurons, genes, proteins) for a large number of configurations. However, the most biologically relevantinformation lies in their connectivity and in the way their components interact, information that these techniques aren't able to record directly. The aim of this thesis is to study statistical methods for inferring information about the connectivity of complex networks starting from experimental data. The subject is approached from a statistical physics point of view drawing from the arsenal of methods developed in the study of spin glasses. Spin-glasses are prototypes of networks of discrete variables interacting in a complex way and are widely used to model biological networks. After an introduction of the models used and a discussion on the biological motivation of the thesis, all known methods of network inference are introduced and analysed from the point of view of their performance. Then, in the third part of the thesis, a new method is proposed which relies in the remark that the interactions in biology are not necessarily symmetric (i.e. the interaction from node A to node B is not the same as the one from B to A). It is shown that this assumption leads to methods that are both exact and efficient. This means that the interactions can be computed exactly, given a sufficient amount of data, and in a reasonable amount of time. This is an important original contribution since no other method is known to be both exact and efficient.

L’entropie d’une distribution sur un ensemble de variables aléatoires discrètes est toujours bornée par l’entropie de la distribution factorisée correspondante. Cette propriété est due à la sous-modularité de l’entropie. Par ailleurs, les fonctions sous-modulaires sont une généralisation des fonctions de rang des matroïdes ; ainsi, les fonctions linéaires sur les polytopes associés peuvent être minimisées exactement par un algorithme glouton. Dans ce manuscrit, nous exploitons ces liens entre les structures des modèles graphiques et les fonctions sous-modulaires. Nous utilisons des algorithmes gloutons pour optimiser des fonctions linéaires sur des polytopes liés aux matroïdes graphiques et hypergraphiques pour apprendre la structure de modèles graphiques, tandis que nous utilisons des algorithmes d’inférence sur les graphes pour optimiser des fonctions sous-modulaires. La première contribution de cette thèse consiste à approcher par maximum de vraisemblance une distribution de probabilité par une distribution factorisable et de complexité algorithmique contrôlée. Comme cette complexité est exponentielle dans la largeur arborescente du graphe, notre but est d’apprendre un graphe décomposable avec une largeur arborescente bornée, ce qui est connu pour être NP-difficile. Nous posons ce problème comme un problème d’optimisation combinatoire et nous proposons une relaxation convexe basée sur les matroïdes graphiques et hypergraphiques. Ceci donne lieu à une solution approchée avec une bonne performance pratique. Pour la seconde contribution principale, nous utilisons le fait que l’entropie d’une distribution est toujours bornée par l’entropie de sa distribution factorisée associée, comme conséquence principale de la sous-modularité, permettant une généralisation à toutes les fonctions sous-modulaires de bornes basées sur les concepts de modèles graphiques. Un algorithme est développé pour maximiser les fonctions sous-modulaires, un autre problème NP-difficile, en maximisant ces bornes en utilisant des algorithmes d’inférence vibrationnels sur les graphes. En troisième contribution, nous proposons et analysons des algorithmes visant à minimiser des fonctions sous-modulaires pouvant s’écrire comme somme de fonctions plus simples. Nos algorithmes n’utilisent que des oracles de ces fonctions simple basés sur minimisation sous-modulaires et de variation totale de telle fonctions. / The entropy of a probability distribution on a set of discrete random variables is always bounded by the entropy of its factorisable counterpart. This is due to the submodularity of entropy on the set of discrete random variables. Submodular functions are also generalisation of matroid rank function; therefore, linear functions may be optimised on the associated polytopes exactly using a greedy algorithm. In this manuscript, we exploit these links between the structures of graphical models and submodular functions: we use greedy algorithms to optimise linear functions on the polytopes related to graphic and hypergraphic matroids for learning the structures of graphical models, while we use inference algorithms on graphs to optimise submodular functions.The first main contribution of the thesis aims at approximating a probabilistic distribution with a factorisable tractable distribution under the maximum likelihood framework. Since the tractability of exact inference is exponential in the treewidth of the decomposable graph, our goal is to learn bounded treewidth decomposable graphs, which is known to be NP-hard. We pose this as a combinatorial optimisation problem and provide convex relaxations based on graphic and hypergraphic matroids. This leads to an approximate solution with good empirical performance. In the second main contribution, we use the fact that the entropy of a probability distribution is always bounded by the entropy of its factorisable counterpart mainly as a consequence of submodularity. This property of entropy is generalised to all submodular functions and bounds based on graphical models are proposed. We refer to them as graph-based bounds. An algorithm is developped to maximise submodular functions, which is NPhard, by maximising the graph-based bound using variational inference algorithms on graphs. As third contribution, we propose and analyse algorithms aiming at minimizing submodular functions that can be written as sum of simple functions. Our algorithms only make use of submodular function minimisation and total variation oracles of simple functions.

Dans un réseau bayésien, une observation sur une variable signifie en général que cette variable est instanciée. Ceci signifie que l’observateur peut affirmer avec certitude que la variable est dans l’état signalé. Cette thèse porte sur d’autres types d’observations, souvent appelées observations incertaines, qui ne peuvent pas être représentées par la simple affectation de la variable. Cette thèse clarifie et étudie les différents concepts d’observations incertaines et propose différentes applications des observations incertaines dans les réseaux bayésiens.Nous commençons par dresser un état des lieux sur les observations incertaines dans les réseaux bayésiens dans la littérature et dans les logiciels, en termes de terminologie, de définition, de spécification et de propagation. Il en ressort que le vocabulaire n'est pas clairement établi et que les définitions proposées couvrent parfois des notions différentes.Nous identifions trois types d’observations incertaines dans les réseaux bayésiens et nous proposons la terminologie suivante : observation de vraisemblance, observation probabiliste fixe et observation probabiliste non-fixe. Nous exposons ensuite la façon dont ces observations peuvent être traitées et propagées.Enfin, nous donnons plusieurs exemples d’utilisation des observations probabilistes fixes dans les réseaux bayésiens. Le premier exemple concerne la propagation d'observations sur une sous-population, appliquée aux systèmes d'information géographique. Le second exemple concerne une organisation de plusieurs agents équipés d'un réseau bayésien local et qui doivent collaborer pour résoudre un problème. Le troisième exemple concerne la prise en compte d'observations sur des variables continues dans un RB discret. Pour cela, l'algorithme BN-IPFP-1 a été implémenté et utilisé sur des données médicales de l'hôpital Bourguiba de Sfax. / In a Bayesian network, evidence on a variable usually signifies that this variable is instantiated, meaning that the observer can affirm with certainty that the variable is in the signaled state. This thesis focuses on other types of evidence, often called uncertain evidence, which cannot be represented by the simple assignment of the variables. This thesis clarifies and studies different concepts of uncertain evidence in a Bayesian network and offers various applications of uncertain evidence in Bayesian networks.Firstly, we present a review of uncertain evidence in Bayesian networks in terms of terminology, definition, specification and propagation. It shows that the vocabulary is not clear and that some terms are used to represent different concepts.We identify three types of uncertain evidence in Bayesian networks and we propose the followingterminology: likelihood evidence, fixed probabilistic evidence and not-fixed probabilistic evidence. We define them and describe updating algorithms for the propagation of uncertain evidence. Finally, we propose several examples of the use of fixed probabilistic evidence in Bayesian networks. The first example concerns evidence on a subpopulation applied in the context of a geographical information system. The second example is an organization of agent encapsulated Bayesian networks that have to collaborate together to solve a problem. The third example concerns the transformation of evidence on continuous variables into fixed probabilistic evidence. The algorithm BN-IPFP-1 has been implemented and used on medical data from CHU Habib Bourguiba in Sfax.

No description available.

Le raisonnement causal est au centre des facultés intellectuelles humaines qui nous permettent de transférer nos connaissances acquises dans des situations très différentes de l'expérience vécue à partir de peu de nouvelles observations. En fait, notre science en entier se base sur l'hypothèse qu'on puisse expliquer tous les phénomènes de l'univers à partir d'un nombre relativement petit de principes simples et constants à travers le temps qui donnent naissance au monde complexe qui nous entoure grâce au très grand nombre de conditions expérimentales possibles, qui correspondent à des interventions dans un modèle causal graphique. La découverte algorithmique de ces mécanismes semble donc être un pilier important, non seulement afin de produire des agents artificiels dotés de capacités cognitives humaines, mais également en vue d'automatiser la découverte scientifique. Nous nous penchons sur une variante du problème de la découverte causale dans laquelle les données observées ne correspondent pas directement aux variables d'intérêt, que l'on considère latentes. Nous utilisons les réseaux de flot génératifs pour apprendre une distribution bayésienne a posteriori définie sur la structure des réseaux bayésiens latents et sur les valeurs des variables latentes. / Causal reasoning is at the center of the human intellectual abilities that allow us to transfer our acquired knowledge in situations that are very different from our past experience from few new observations. In fact, our whole science is based on the assumption that we can explain all the phenomena of the universe from a relatively small set of simple principles that are constant through time and that give rise to the complex world surrounding us due to the very large number of possible experimental conditions that correspond to interventions in a causal graphical model. The algorithmic discovery of these mechanisms thus seems to be an important pillar, not only to create artificial agents endowed with human cognitive abilities, but also to automate scientific discovery. We are looking into a variant of the causal discovery problem in which the observed data does not directly correspond to the variables of interest, which we consider to be latent. We use Generative Flow Networks to learn a Bayesian posterior distribution defined over latent Bayesian networks and over the values of the latent variables.

Sequence alignment is one of the most intensely studied problems in bioinformatics, and is an important step in a wide range of analyses. An issue that has gained much attention in recent years is the fact that downstream analyses are often highly sensitive to the specific choice of alignment. One way to address this is to jointly sample alignments along with other parameters of interest. In order to extend the range of applicability of this approach, the first chapter of this thesis introduces a probabilistic evolutionary model for protein structures on a phylogenetic tree; since protein structures typically diverge much more slowly than sequences, this allows for more reliable detection of remote homologies, improving the accuracy of the resulting alignments and trees, and reducing sensitivity of the results to the choice of dataset. In order to carry out inference under such a model, a number of new Markov chain Monte Carlo approaches are developed, allowing for more efficient convergence and mixing on the high-dimensional parameter space. The second part of the thesis presents a directed acyclic graph (DAG)-based approach for representing a collection of sampled alignments. This DAG representation allows the initial collection of samples to be used to generate a larger set of alignments under the same approximate distribution, enabling posterior alignment probabilities to be estimated reliably from a reasonable number of samples. If desired, summary alignments can then be generated as maximum-weight paths through the DAG, under various types of loss or scoring functions. The acyclic nature of the graph also permits various other types of algorithms to be easily adapted to operate on the entire set of alignments in the DAG. In the final part of this work, methodology is introduced for alignment-DAG-based sequence annotation using hidden Markov models, and RNA secondary structure prediction using stochastic context-free grammars. Results on test datasets indicate that the additional information contained within the DAG allows for improved predictions, resulting in substantial gains over simply analysing a set of alignments one by one.

Des techniques expérimentales récentes ont donné la possibilité d'acquérir un très grand nombre de données concernant des réseaux biologiques complexes, comme des réseaux de neurones, des réseaux de gènes et des réseaux d'interactions de protéines. Ces techniques sont capables d'enregistrer les états des composantes individuelles de ces réseaux (neurones, gènes, protéines) pour un grand nombre de configurations. Cependant, l'information la plus pertinente biologiquement se trouve dans la connectivité de ces systèmes et dans la façon précise avec laquelle ces composantes interagissent, information que les techniques expérimentales ne sont pas au point d'observer directement. Le bût de cette thèse est d'étudier les méthodes statistiques nécessaires pour inférer de l'information sur la connectivité des réseaux complexes en partant des données expérimentales. Ce sujet est traité par le point de vue de la physique statistique, en puisant de l'arsenal de méthodes théoriques qui ont été développées pour l'étude des verres de spins. Les verres de spins sont des exemples de réseaux à variables discrètes qui interagissent de façon complexe et sont souvent utilisés pour modéliser des réseaux biologiques. Après une introduction sur les modèles utilisés ainsi qu'une discussion sur la motivation biologique de cette thèse, toutes les méthodes d'inférence de réseaux connues sont présentées et analysées du point de vue de leur performance. Par la suite, dans la troisième partie de la thèse, un nouvelle méthode est proposée qui s'appuie sur la remarque que les interactions en biologie ne sont pas nécessairement symétriques (c'est-à-dire l'interaction entre les noeuds A et B n'est pas la même dans les deux directions). Il est démontré que cette assomption conduit à des méthodes qui sont capables de prédire les interactions de façon exacte, étant donné un nombre suffisant de données, tout en utilisant un temps de calcul polynomial. Ceci est un résultat original important car toutes les autres méthodes connues sont soit exactes et non-polynomiales soit inexactes et polynomiales. / Recent experimental techniques in biology made possible the acquisition of overwhelming amounts of data concerning complex biological networks, such as neural networks, gene regulation networks and protein-protein interaction networks. These techniques are able to record states of individual components of such networks (neurons, genes, proteins) for a large number of configurations. However, the most biologically relevantinformation lies in their connectivity and in the way their components interact, information that these techniques aren't able to record directly. The aim of this thesis is to study statistical methods for inferring information about the connectivity of complex networks starting from experimental data. The subject is approached from a statistical physics point of view drawing from the arsenal of methods developed in the study of spin glasses. Spin-glasses are prototypes of networks of discrete variables interacting in a complex way and are widely used to model biological networks. After an introduction of the models used and a discussion on the biological motivation of the thesis, all known methods of network inference are introduced and analysed from the point of view of their performance. Then, in the third part of the thesis, a new method is proposed which relies in the remark that the interactions in biology are not necessarily symmetric (i.e. the interaction from node A to node B is not the same as the one from B to A). It is shown that this assumption leads to methods that are both exact and efficient. This means that the interactions can be computed exactly, given a sufficient amount of data, and in a reasonable amount of time. This is an important original contribution since no other method is known to be both exact and efficient.

Motivation and background The enormous amount of capabilities that every human learns throughout his life, is probably among the most remarkable and fascinating aspects of life. Learning has therefore drawn lots of interest from scientists working in very different fields like philosophy, biology, sociology, educational sciences, computer sciences and mathematics. This thesis focuses on the information theoretical and mathematical aspects of learning. We are interested in the learning process of an agent (which can be for example a human, an animal, a robot, an economical institution or a state) that interacts with its environment. Common models for this interaction are Markov decision processes (MDPs) and partially observable Markov decision processes (POMDPs). Learning is then considered to be the maximization of the expectation of a predefined reward function. In order to formulate general principles (like a formal definition of curiosity-driven learning or avoidance of unpleasant situation) in a rigorous way, it might be desirable to have a theoretical framework for the optimization of more complex functionals of the underlying process law. This might include the entropy of certain sensor values or their mutual information. An optimization of the latter quantity (also known as predictive information) has been investigated intensively both theoretically and experimentally using computer simulations by N. Ay, R. Der, K Zahedi and G. Martius. In this thesis, we develop a mathematical theory for learning in the sensorimotor loop beyond expected reward maximization. Approaches and results This thesis covers four different topics related to the theory of learning in the sensorimotor loop. First of all, we need to specify the model of an agent interacting with the environment, either with learning or without learning. This interaction naturally results in complex causal dependencies. Since we are interested in asymptotic properties of learning algorithms, it is necessary to consider infinite time horizons. It turns out that the well-understood theory of causal networks known from the machine learning literature is not powerful enough for our purpose. Therefore we extend important theorems on causal networks to infinite graphs and general state spaces using analytical methods from measure theoretic probability theory and the theory of discrete time stochastic processes. Furthermore, we prove a generalization of the strong Markov property from Markov processes to infinite causal networks. Secondly, we develop a new idea for a projected stochastic constraint optimization algorithm. Generally a discrete gradient ascent algorithm can be used to generate an iterative sequence that converges to the stationary points of a given optimization problem. Whenever the optimization takes place over a compact subset of a vector space, it is possible that the iterative sequence leaves the constraint set. One possibility to cope with this problem is to project all points to the constraint set using Euclidean best-approximation. The latter is sometimes difficult to calculate. A concrete example is an optimization over the unit ball in a matrix space equipped with operator norm. Our idea consists of a back-projection using quasi-projectors different from the Euclidean best-approximation. In the matrix example, there is another canonical way to force the iterative sequence to stay in the constraint set: Whenever a point leaves the unit ball, it is divided by its norm. For a given target function, this procedure might introduce spurious stationary points on the boundary. We show that this problem can be circumvented by using a gradient that is tailored to the quasi-projector used for back-projection. We state a general technical compatibility condition between a quasi-projector and a metric used for gradient ascent, prove convergence of stochastic iterative sequences and provide an appropriate metric for the unit-ball example. Thirdly, a class of learning problems in the sensorimotor loop is defined and motivated. This class of problems is more general than the usual expected reward maximization and is illustrated by numerous examples (like expected reward maximization, maximization of the predictive information, maximization of the entropy and minimization of the variance of a given reward function). We also provide stationarity conditions together with appropriate gradient formulas. Last but not least, we prove convergence of a stochastic optimization algorithm (as considered in the second topic) applied to a general learning problem (as considered in the third topic). It is shown that the learning algorithm converges to the set of stationary points. Among others, the proof covers the convergence of an improved version of an algorithm for the maximization of the predictive information as proposed by N. Ay, R. Der and K. Zahedi. We also investigate an application to a linear Gaussian dynamic, where the policies are encoded by the unit-ball in a space of matrices equipped with operator norm.