Face au déluge actuel de données principalement non structurées, les graphes ont démontré, dans une variété de domaines scientifiques, leur importance croissante comme language abstrait pour décrire des interactions complexes entre des objets complexes. L’un des principaux défis posés par l’étude de ces réseaux est l’inférence de propriétés macroscopiques à grande échelle, affectant un grand nombre d’objets ou d’agents, sur la seule base des interactions microscopiquesqu’entretiennent leurs constituants élémentaires. La physique statistique, créée précisément dans le but d’obtenir les lois macroscopiques de la thermodynamique à partir d’un modèle idéal de particules en interaction, fournit une intuition décisive dans l’étude des réseaux complexes.Dans cette thèse, nous utilisons des méthodes issues de la physique statistique des systèmes désordonnés pour mettre au point et analyser de nouveaux algorithmes d’inférence sur les graphes. Nous nous concentrons sur les méthodes spectrales, utilisant certains vecteurs propres de matrices bien choisies, et sur les graphes parcimonieux, qui contiennent une faible quantité d’information. Nous développons une théorie originale de l’inférence spectrale, fondée sur une relaxation de l’optimisation de certaines énergies libres en champ moyen. Notre approche est donc entièrement probabiliste, et diffère considérablement des motivations plus classiques fondées sur l’optimisation d’une fonction de coût. Nous illustrons l’efficacité de notre approchesur différents problèmes, dont la détection de communautés, la classification non supervisée à partir de similarités mesurées aléatoirement, et la complétion de matrices. / In an era of unprecedented deluge of (mostly unstructured) data, graphs are proving more and more useful, across the sciences, as a flexible abstraction to capture complex relationships between complex objects. One of the main challenges arising in the study of such networks is the inference of macroscopic, large-scale properties affecting a large number of objects, based solely on he microscopic interactions between their elementary constituents. Statistical physics, precisely created to recover the macroscopic laws of thermodynamics from an idealized model of interacting particles, provides significant insight to tackle such complex networks.In this dissertation, we use methods derived from the statistical physics of disordered systems to design and study new algorithms for inference on graphs. Our focus is on spectral methods, based on certain eigenvectors of carefully chosen matrices, and sparse graphs, containing only a small amount of information. We develop an original theory of spectral inference based on a relaxation of various meanfield free energy optimizations. Our approach is therefore fully probabilistic, and contrasts with more traditional motivations based on the optimization of a cost function. We illustrate the efficiency of our approach on various problems, including community detection, randomized similarity-based clustering, and matrix completion.

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.

<p>Graphs are a natural choice for understanding the associations between variables, and assuming a probabilistic embedding for the graph structure leads to a variety of graphical models that enable us to understand these associations even further. In the realm of high-dimensional data, where the number of associations between interacting variables is far greater than the available number of data points, the goal is to infer a sparse graph. In this thesis, we make contributions in the domain of Bayesian graphical models, where our prior belief on the graph structure, encoded via uncertainty on the model parameters, enables the estimation of sparse graphs.</p> <p><br></p> <p>We begin with the Gaussian Graphical Model (GGM) in Chapter 2, one of the simplest and most famous graphical models, where the joint distribution of interacting variables is assumed to be Gaussian. In GGMs, the conditional independence among variables is encoded in the inverse of the covariance matrix, also known as the precision matrix. Under a Bayesian framework, we propose a novel prior--penalty dual called the `graphical horseshoe-like' prior and penalty, to estimate precision matrix. We also establish the posterior convergence of the precision matrix estimate and the frequentist consistency of the maximum a posteriori (MAP) estimator.</p> <p><br></p> <p>In Chapter 3, we develop a general framework based on local linear approximation for MAP estimation of the precision matrix in GGMs. This general framework holds true for any graphical prior, where the element-wise priors can be written as a Laplace scale mixture. As an application of the framework, we perform MAP estimation of the precision matrix under the graphical horseshoe penalty.</p> <p><br></p> <p>In Chapter 4, we focus on graphical models where the joint distribution of interacting variables cannot be assumed Gaussian. Motivated by the quantile graphical models, where the Gaussian likelihood assumption is relaxed, we draw inspiration from the domain of precision medicine, where personalized inference is crucial to tailor individual-specific treatment plans. With an aim to infer Directed Acyclic Graphs (DAGs), we propose a novel quantile DAG learning framework, where the DAGs depend on individual-specific covariates, making personalized inference possible. We demonstrate the potential of this framework in the regime of precision medicine by applying it to infer protein-protein interaction networks in Lung adenocarcinoma and Lung squamous cell carcinoma.</p> <p><br></p> <p>Finally, we conclude this thesis in Chapter 5, by developing a novel framework to compute the marginal likelihood in a GGM, addressing a longstanding open problem. Under this framework, we can compute the marginal likelihood for a broad class of priors on the precision matrix, where the element-wise priors on the diagonal entries can be written as gamma or scale mixtures of gamma random variables and those on the off-diagonal terms can be represented as normal or scale mixtures of normal. This result paves new roads for model selection using Bayes factors and tuning of prior hyper-parameters.</p>

L'interaction face-à-face représente une des formes les plus fondamentales de la communication humaine. C'est un système dynamique multimodal et couplé – impliquant non seulement la parole mais de nombreux segments du corps dont le regard, l'orientation de la tête, du buste et du corps, les gestes faciaux et brachio-manuels, etc – d'une grande complexité. La compréhension et la modélisation de ce type de communication est une étape cruciale dans le processus de la conception des agents interactifs capables d'engager des conversations crédibles avec des partenaires humains. Concrètement, un modèle de comportement multimodal destiné aux agents sociaux interactifs fait face à la tâche complexe de générer un comportement multimodal étant donné une analyse de la scène et une estimation incrémentale des objectifs conjoints visés au cours de la conversation. L'objectif de cette thèse est de développer des modèles de comportement multimodal pour permettre aux agents artificiels de mener une communication co-verbale pertinente avec un partenaire humain. Alors que l'immense majorité des travaux dans le domaine de l'interaction humain-agent repose essentiellement sur des modèles à base de règles, notre approche se base sur la modélisation statistique des interactions sociales à partir de traces collectées lors d'interactions exemplaires, démontrées par des tuteurs humains. Dans ce cadre, nous introduisons des modèles de comportement dits "sensori-moteurs", qui permettent à la fois la reconnaissance des états cognitifs conjoints et la génération des signaux sociaux d'une manière incrémentale. En particulier, les modèles de comportement proposés ont pour objectif d'estimer l'unité d'interaction (IU) dans laquelle sont engagés de manière conjointe les interlocuteurs et de générer le comportement co-verbal du tuteur humain étant donné le comportement observé de son/ses interlocuteur(s). Les modèles proposés sont principalement des modèles probabilistes graphiques qui se basent sur les chaînes de markov cachés (HMM) et les réseaux bayésiens dynamiques (DBN). Les modèles ont été appris et évalués – notamment comparés à des classifieurs classiques – sur des jeux de données collectés lors de deux différentes interactions face-à-face. Les deux interactions ont été soigneusement conçues de manière à collecter, en un minimum de temps, un nombre suffisant d'exemplaires de gestion de l'attention mutuelle et de deixis multimodale d'objets et de lieux. Nos contributions sont complétées par des méthodes originales d'interprétation et d'évaluation des propriétés des modèles proposés. En comparant tous les modèles avec les vraies traces d'interactions, les résultats montrent que le modèle HMM, grâce à ses propriétés de modélisation séquentielle, dépasse les simples classifieurs en terme de performances. Les modèles semi-markoviens (HSMM) ont été également testé et ont abouti à un meilleur bouclage sensori-moteur grâce à leurs propriétés de modélisation des durées des états. Enfin, grâce à une structure de dépendances riche apprise à partir des données, le modèle DBN a les performances les plus probantes et démontre en outre la coordination multimodale la plus fidèle aux évènements multimodaux originaux. / Face to face interaction is one of the most fundamental forms of human communication. It is a complex multimodal and coupled dynamic system involving not only speech but of numerous segments of the body among which gaze, the orientation of the head, the chest and the body, the facial and brachiomanual movements, etc. The understanding and the modeling of this type of communication is a crucial stage for designing interactive agents capable of committing (hiring) credible conversations with human partners. Concretely, a model of multimodal behavior for interactive social agents faces with the complex task of generating gestural scores given an analysis of the scene and an incremental estimation of the joint objectives aimed during the conversation. The objective of this thesis is to develop models of multimodal behavior that allow artificial agents to engage into a relevant co-verbal communication with a human partner. While the immense majority of the works in the field of human-agent interaction (HAI) is scripted using ruled-based models, our approach relies on the training of statistical models from tracks collected during exemplary interactions, demonstrated by human trainers. In this context, we introduce "sensorimotor" models of behavior, which perform at the same time the recognition of joint cognitive states and the generation of the social signals in an incremental way. In particular, the proposed models of behavior have to estimate the current unit of interaction ( IU) in which the interlocutors are jointly committed and to predict the co-verbal behavior of its human trainer given the behavior of the interlocutor(s). The proposed models are all graphical models, i.e. Hidden Markov Models (HMM) and Dynamic Bayesian Networks (DBN). The models were trained and evaluated - in particular compared with classic classifiers - using datasets collected during two different interactions. Both interactions were carefully designed so as to collect, in a minimum amount of time, a sufficient number of exemplars of mutual attention and multimodal deixis of objects and places. Our contributions are completed by original methods for the interpretation and comparative evaluation of the properties of the proposed models. By comparing the output of the models with the original scores, we show that the HMM, thanks to its properties of sequential modeling, outperforms the simple classifiers in term of performances. The semi-Markovian models (HSMM) further improves the estimation of sensorimotor states thanks to duration modeling. Finally, thanks to a rich structure of dependency between variables learnt from the data, the DBN has the most convincing performances and demonstrates both the best performance and the most faithful multimodal coordination to the original multimodal events.