Contributions à l'apprentissage statistique de lois a posteriori pour l'inférence bayésienne sans vraisemblance
Elouan Argouarc h · theses.fr (ABES) · 2024
Bayesian posterior inference is used in many scientific applications and is a prevalent methodology for decision-making under uncertainty. It enables practitioners to confront real-world observations with relevant observation models, and in turn, infer the distribution over an explanatory variable. In many fields and practical applications, we consider ever more intricate observation models for their otherwise scientific relevance, but at the cost of intractable probability density functions. As a result, both the likelihood and the posterior are unavailable, making posterior inference using the usual Monte Carlo methods unfeasible.In this thesis, we suppose that the observation model provides a recorded dataset, and our aim is to bring together Bayesian inference and statistical learning methods to perform posterior inference in a likelihood-free setting. This problem, formulated as learning an approximation of a posterior distribution, includes the usual statistical learning tasks of regression and classification modeling, but it can also be an alternative to Approximate Bayesian Computation methods in the context of simulation-based inference, where the observation model is instead a simulation model with implicit density.The aim of this thesis is to propose methodological contributions for Bayesian posterior learning. More precisely, our main goal is to compare different learning methods under the scope of Monte Carlo sampling and uncertainty quantification.We first consider the posterior approximation based on the likelihood-to-evidence ratio, which has the main advantage that it turns a problem of conditional density learning into a problem of binary classification. In the context of Monte Carlo sampling, we propose a methodology for sampling from such a posterior approximation. We leverage the structure of the underlying model, which is conveniently compatible with the usual ratio-based sampling algorithms, to obtain straightforward, parameter-free, and density-free sampling procedures.We then turn to the problem of uncertainty quantification. On the one hand, normalized models such as the discriminative construction are easy to apply in the context of Bayesian uncertainty quantification. On the other hand, while unnormalized models, such as the likelihood-to-evidence-ratio, are not easily applied in uncertainty-aware learning tasks, a specific unnormalized construction, which we refer to as generative, is indeed compatible with Bayesian uncertainty quantification via the posterior predictive distribution. In this context, we explain how to carry out uncertainty quantification in both modeling techniques, and we then propose a comparison of the two constructions under the scope of Bayesian learning.We finally turn to the problem of parametric modeling with tractable density, which is indeed a requirement for epistemic uncertainty quantification in generative and discriminative modeling methods. We propose a new construction of a parametric model, which is an extension of both mixture models and normalizing flows. This model can be applied to many different types of statistical problems, such as variational inference, density estimation, and conditional density estimation, as it benefits from rapid and exact density evaluation, a straightforward sampling scheme, and a gradient reparameterization approach.