The extension of the CJS model to include individual, continuous, time-varying covariates relies on the estimation of covariate values on occasions on which individuals were not captured. Fitting this model in a Bayesian framework typically involves the implementation of a Markov chain Monte Carlo (MCMC) algorithm, such as a Gibbs sampler, to sample from the posterior distribution. For large data sets with many missing covariate values that must be estimated, this creates a computational issue, as each iteration of the MCMC algorithm requires sampling from the full conditional distributions of each missing covariate value. This dissertation examines two solutions to address this problem. First, I explore variational Bayesian algorithms, which derive inference from an approximation to the posterior distribution that can be fit quickly in many complex problems. Second, I consider an alternative approximation to the posterior distribution derived by truncating the individual capture histories in order to reduce the number of missing covariates that must be updated during the MCMC sampling algorithm. In both cases, the increased computational efficiency comes at the cost of producing approximate inferences. The variational Bayesian algorithms generally do not estimate the posterior variance very accurately and do not directly address the issues with estimating many missing covariate values. Meanwhile, the truncated CJS model provides a more significant improvement in computational efficiency while inflating the posterior variance as a result of discarding some of the data. Both approaches are evaluated via simulation studies and a large mark-recapture data set consisting of cliff swallow weights and capture histories.

京都大学 / 0048 / 新制・課程博士 / 博士(情報学) / 甲第21208号 / 情博第661号 / 新制||情||114(附属図書館) / 京都大学大学院情報学研究科知能情報学専攻 / (主査)教授 美濃 導彦, 教授 河原 達也, 教授 中村 裕一 / 学位規則第4条第1項該当 / Doctor of Informatics / Kyoto University / DFAM

The cross-situational word learning paradigm argues that word meanings can be approximated by word-object associations, computed from co-occurrence statistics between words and entities in the world. Lexicon acquisition involves simultaneously guessing (1) which objects are being talked about (the ”meaning”) and (2) which words relate to those objects. However, most modeling work focuses on acquiring meanings for isolated words, largely neglecting relationships between words or physical entities, which can play an important role in learning. Semantic parsing, on the other hand, aims to learn a mapping between entire utterances and compositional meaning representations where such relations are central. The focus is the mapping between meaning and words, while utterance meanings are treated as observed quantities. Here, we extend the joint inference problem of word learning to account for compositional meanings by incorporating a semantic parsing model for relating utterances to non-linguistic context. Integrating semantic parsing and word learning permits us to explore the impact of word-word and concept-concept relations. The result is a joint-inference problem inherited from the word learning setting where we must simultaneously learn utterance-level and individual word meanings, only now we also contend with the many possible relationships between concepts in the meaning and words in the sentence. To simplify design, we factorize the model into separate modules, one for each of the world, the meaning, and the words, and merge them into a single synchronous grammar for joint inference. There are three main contributions. First, we introduce a novel word learning model and accompanying semantic parser. Second, we produce a corpus which allows us to demonstrate the importance of structure in word learning. Finally, we also present a number of technical innovations required for implementing such a model.

Nous traitons le sujet de l’estimation conjointe des signaux aléatoires dépendant de paramètres déterministes et inconnus. Premièrement, on aborde le sujet du côté applicatif en proposant deux algorithmes de séparation de la parole voisée mono-capteur. Dans le premier, nous utilisons le modèle autorégressif de la parole qui décrit les corrélations court et long termes (quasi-périodique) pour formuler un modèle d’état dépendant de paramètres inconnus. EM-Kalman est ainsi utilisé pour estimer conjointement les sources et les paramètres. Dans le deuxième, nous proposons une méthode fréquentielle pour le même modèle de la parole où les sources et les paramètres sont estimés séparément. Les observations sont découpées à l’aide d’un fenêtrage bien conçu pour assurer une reconstruction parfaite des sources après. Les paramètres (de l’enveloppe spectrale) sont estimés en maximisant le critère du GML exprimé avec la matrice de covariance paramétrée que nous modélisons plus correctement en tenant compte de l’effet du fenêtrage. Le filtre de Wiener est utilisé pour estimer les sources. Deuxièmement, on aborde l’estimation conjointe d’un point de vue plus théorique en s'interrogeant sur les performances relatives de l’estimation conjointe par rapport à l’estimation séparée d’une manière générale. Nous considérons le cas conjointement Gaussien (observations et variables cachées) et trois méthodes itératives d'estimation conjointe: MAP en alternance avec ML, biaisé même asymptotiquement pour les paramètres, EM qui converge asymptotiquement vers ML et VB que nous prouvons converger asymptotiquement vers la solution ML pour les paramètres déterministes. / The thesis is composed of two parts. In the first part, we deal with the monaural speech separation problem. We propose two algorithms. In the first algorithm, we exploit the joint autoregressive model that models short and long (periodic) correlations of Gaussian speech signals to formulate a state space model with unknown parameters. The EM-Kalman algorithm is then used to estimate jointly the sources (involved in the state vector) and the parameters of the model. In the second algorithm, we use the same speech model but this time in the frequency domain (quasi-periodic Gaussian sources with AR spectral envelope). Observation data is sliced using a well-designed window. Parameters are estimated separately from the sources by optimizing the Gaussian ML criterion expressed using the sample and parameterized covariance matrices. Classical frequency domain asymptotic methods replace linear convolution by circulant convolution leading to approximation errors. We show how the introduction of windows can lead to slightly more complex frequency domain techniques, replacing diagonal covariance matrices by banded covariance matrices, but with controlled approximation error. The sources are then estimated using the Wiener filtering. The second part is about the relative performance of joint vs. marginalized parameter estimation. We consider jointly Gaussian latent data and observations. We provide contributions to Cramer-Rao bounds, then, we investigate three iterative joint estimation approaches: Alternating MAP/ML which suffers from inconsistent parameter bias, EM which converges to ML and VB that we prove converges asymptotically to the ML solution for parameter estimation.

No description available.

Les approches bayésiennes sont largement utilisées dans le domaine du traitement du signal. Elles utilisent des informations a priori sur les paramètres inconnus à estimer ainsi que des informations sur les observations, pour construire des estimateurs. L'estimateur optimal au sens du coût quadratique est l'un des estimateurs les plus couramment employés. Toutefois, comme la loi a posteriori exacte a très souvent une forme complexe, il faut généralement recourir à des outils d'approximation bayésiens pour l'approcher. Dans ce travail, nous nous intéressons particulièrement à deux types de méthodes: les algorithmes d'échantillonnage Monte Carlo par chaînes de Markov (MCMC) et les approches basées sur des approximations bayésiennes variationnelles (VBA).La thèse est composée de deux parties. La première partie concerne les algorithmes d'échantillonnage. Dans un premier temps, une attention particulière est consacrée à l'amélioration des méthodes MCMC basées sur la discrétisation de la diffusion de Langevin. Nous proposons une nouvelle méthode pour régler la composante directionnelle de tels algorithmes en utilisant une stratégie de Majoration-Minimisation ayant des propriétés de convergence garanties. Les résultats expérimentaux obtenus lors de la restauration d'un signal parcimonieux confirment la rapidité de cette nouvelle approche par rapport à l'échantillonneur usuel de Langevin. Dans un second temps, une nouvelle méthode d'échantillonnage basée sur une stratégie d'augmentation des données est proposée pour améliorer la vitesse de convergence et les propriétés de mélange des algorithmes d'échantillonnage standards. L'application de notre méthode à différents exemples en traitement d'images montre sa capacité à surmonter les difficultés liées à la présence de corrélations hétérogènes entre les coefficients du signal.Dans la seconde partie de la thèse, nous proposons de recourir aux techniques VBA pour la restauration de signaux dégradés par un bruit non-gaussien. Afin de contourner les difficultés liées à la forme compliquée de la loi a posteriori, une stratégie de majoration est employée pour approximer la vraisemblance des données ainsi que la densité de la loi a priori. Grâce à sa flexibilité, notre méthode peut être appliquée à une large classe de modèles et permet d'estimer le signal d'intérêt conjointement au paramètre de régularisation associé à la loi a priori. L'application de cette approche sur des exemples de déconvolution d'images en présence d'un bruit mixte Poisson-gaussien, confirme ses bonnes performances par rapport à des méthodes supervisées de l'état de l'art. / Bayesian approaches are widely used in signal processing applications. In order to derive plausible estimates of original parameters from their distorted observations, they rely on the posterior distribution that incorporates prior knowledge about the unknown parameters as well as informations about the observations. The posterior mean estimator is one of the most commonly used inference rule. However, as the exact posterior distribution is very often intractable, one has to resort to some Bayesian approximation tools to approximate it. In this work, we are mainly interested in two particular Bayesian methods, namely Markov Chain Monte Carlo (MCMC) sampling algorithms and Variational Bayes approximations (VBA).This thesis is made of two parts. The first one is dedicated to sampling algorithms. First, a special attention is devoted to the improvement of MCMC methods based on the discretization of the Langevin diffusion. We propose a novel method for tuning the directional component of such algorithms using a Majorization-Minimization strategy with guaranteed convergence properties.Experimental results on the restoration of a sparse signal confirm the performance of this new approach compared with the standard Langevin sampler. Second, a new sampling algorithm based on a Data Augmentation strategy, is proposed to improve the convergence speed and the mixing properties of standard MCMC sampling algorithms. Our methodological contributions are validated on various applications in image processing showing the great potentiality of the proposed method to manage problems with heterogeneous correlations between the signal coefficients.In the second part, we propose to resort to VBA techniques to build a fast estimation algorithm for restoring signals corrupted with non-Gaussian noise. In order to circumvent the difficulties raised by the intricate form of the true posterior distribution, a majorization technique is employed to approximate either the data fidelity term or the prior density. Thanks to its flexibility, the proposed approach can be applied to a broad range of data fidelity terms allowing us to estimate the target signal jointly with the associated regularization parameter. Illustration of this approach through examples of image deconvolution in the presence of mixed Poisson-Gaussian noise, show the good performance of the proposed algorithm compared with state of the art supervised methods.

<div>Markov jump processes are continuous-time stochastic processes widely used in a variety of applied disciplines. Inference typically proceeds via Markov chain Monte Carlo (MCMC), the state-of-the-art being a uniformization-based auxiliary variable Gibbs sampler. This was designed for situations where the process parameters are known, and Bayesian inference over unknown parameters is typically carried out by incorporating it into a larger Gibbs sampler. This strategy of sampling parameters given path, and path given parameters can result in poor Markov chain mixing.</div><div><br></div><div>In this thesis, we focus on the problem of path and parameter inference for Markov jump processes.</div><div><br></div><div>In the first part of the thesis, a simple and efficient MCMC algorithm is proposed to address the problem of path and parameter inference for Markov jump processes. Our scheme brings Metropolis-Hastings approaches for discrete-time hidden Markov models to the continuous-time setting, resulting in a complete and clean recipe for parameter and path inference in Markov jump processes. In our experiments, we demonstrate superior performance over Gibbs sampling, a more naive Metropolis-Hastings algorithm we propose, as well as another popular approach, particle Markov chain Monte Carlo. We also show our sampler inherits geometric mixing from an ‘ideal’ sampler that is computationally much more expensive.</div><div><br></div><div>In the second part of the thesis, a novel collapsed variational inference algorithm is proposed. Our variational inference algorithm leverages ideas from discrete-time Markov chains, and exploits a connection between Markov jump processes and discrete-time Markov chains through uniformization. Our algorithm proceeds by marginalizing out the parameters of the Markov jump process, and then approximating the distribution over the trajectory with a factored distribution over segments of a piecewise-constant function. Unlike MCMC schemes that marginalize out transition times of a piecewise-constant process, our scheme optimizes the discretization of time, resulting in significant computational savings. We apply our ideas to synthetic data as well as a dataset of check-in recordings, where we demonstrate superior performance over state-of-the-art MCMC methods.</div><div><br></div>

The ability to recognize handshapes in signing video is essential in algorithms for sign recognition and retrieval. Handshape recognition from isolated images is, however, an insufficiently constrained problem. Many handshapes share similar 3D configurations and are indistinguishable for some hand orientations in 2D image projections. Additionally, significant differences in handshape appearance are induced by the articulated structure of the hand and variants produced by different signers. Linguistic rules involved in the production of signs impose strong constraints on the articulations of the hands, yet, little attention has been paid towards exploiting these constraints in previous works on sign recognition. Among the different classes of signs in any signed language, lexical signs constitute the prevalent class. Morphemes (or, meaningful units) for signs in this class involve a combination of particular handshapes, palm orientations, locations for articulation, and movement type. These are thus analyzed by many sign linguists as analogues of phonemes in spoken languages. Phonological constraints govern the ways in which phonemes combine in American Sign Language (ASL), as in other signed and spoken languages; utilizing these constraints for handshape recognition in ASL is the focus of the proposed thesis. Handshapes in monomorphemic lexical signs are specified at the start and end of the sign. The handshape transition within a sign are constrained to involve either closing or opening of the hand (i.e., constrained to exclusively use either folding or unfolding of the palm and one or more fingers). Furthermore, akin to allophonic variations in spoken languages, both inter- and intra- signer variations in the production of specific handshapes are observed. We propose a Bayesian network formulation to exploit handshape co-occurrence constraints also utilizing information about allophonic variations to aid in handshape recognition. We propose a fast non-rigid image alignment method to gain improved robustness to handshape appearance variations during computation of observation likelihoods in the Bayesian network. We evaluate our handshape recognition approach on a large dataset of monomorphemic lexical signs. We demonstrate that leveraging linguistic constraints on handshapes results in improved handshape recognition accuracy. As part of the overall project, we are collecting and preparing for dissemination a large corpus (three thousand signs from three native signers) of ASL video annotated with linguistic information such as glosses, morphological properties and variations, and start/end handshapes associated with each ASL sign.

<div>Stochastic programs are standard models for decision-making under uncertainty and have been extensively studied in the operations research literature. In general, stochastic programming involves minimizing an expected cost function, where the expectation is with respect to fully specified stochastic models that quantify the aleatoric or `inherent' uncertainty in the decision-making problem. In practice, however, the stochastic models are unknown but can be estimated from data, introducing an additional epistemic uncertainty into the decision-making problem. The Bayesian framework provides a coherent way to quantify the epistemic uncertainty through the posterior distribution by combining prior beliefs of the decision-makers with the observed data. Bayesian methods have been used for data-driven decision-making in various applications such as inventory management, portfolio design, machine learning, optimal scheduling, and staffing, etc.</div><div> </div><div>Bayesian methods are challenging to implement, mainly due to the fact that the posterior is computationally intractable, necessitating the computation of approximate posteriors. Broadly speaking, there are two methods in the literature implementing approximate posterior inference. First are sampling-based methods such as Markov Chain Monte Carlo. Sampling-based methods are theoretically well understood, but they suffer from various issues like high variance, poor scalability to high-dimensional problems, and have complex diagnostics. Consequently, we propose to use optimization-based methods collectively known as variational inference (VI) that use information projections to compute an approximation to the posterior. Empirical studies have shown that VI methods are computationally faster and easily scalable to higher-dimensional problems and large datasets. However, the theoretical guarantees of these methods are not well understood. Moreover, VI methods are empirically and theoretically less explored in the decision-theoretic setting.</div><div><br></div><div> In this thesis, we first propose a novel VI framework for risk-sensitive data-driven decision-making, which we call risk-sensitive variational Bayes (RSVB). In RSVB, we jointly compute a risk-sensitive approximation to the `true' posterior and the optimal decision by solving a minimax optimization problem. The RSVB framework includes the naive approach of first computing a VI approximation to the true posterior and then using it in place of the true posterior for decision-making. We show that the RSVB approximate posterior and the corresponding optimal value and decision rules are asymptotically consistent, and we also compute their rate of convergence. We illustrate our theoretical findings in both parametric as well as nonparametric setting with the help of three examples: the single and multi-product newsvendor model and Gaussian process classification. Second, we present the Bayesian joint chance-constrained stochastic program (BJCCP) for modeling decision-making problems with epistemically uncertain constraints. We discover that using VI methods for posterior approximation can ensure the convexity of the feasible set in (BJCCP) unlike any sampling-based methods and thus propose a VI approximation for (BJCCP). We also show that the optimal value computed using the VI approximation of (BJCCP) are statistically consistent. Moreover, we derive the rate of convergence of the optimal value and compute the rate at which a VI approximate solution of (BJCCP) is feasible under the true constraints. We demonstrate the utility of our approach on an optimal staffing problem for an M/M/c queue. Finally, this thesis also contributes to the growing literature in understanding statistical performance of VI methods. In particular, we establish the frequentist consistency of an approximate posterior computed using a well known VI method that computes an approximation to the posterior distribution by minimizing the Renyi divergence from the ‘true’ posterior.</div>

Single molecule experiments study the kinetics of molecular biological systems. Many such studies generate data that can be described by aggregated hidden Markov models, whereby there is a need of doing inference on such data and models. In this study, model selection in aggregated Hidden Markov models was performed with a criterion of maximum Bayesian evidence. Variational Bayes inference was seen to underestimate the evidence for aggregated model fits. Estimation of the evidence integral by brute force Monte Carlo integration theoretically always converges to the correct value, but it converges in far from tractable time. Nested sampling is a promising method for solving this problem by doing faster Monte Carlo integration, but it was here seen to have difficulties generating uncorrelated samples.