Algorithms for Approximate Bayesian Inference with Applications to Astronomical Data Analysis

Bayesian inference is a theoretically well-founded and conceptually simple approach to data analysis. The computations in practical problems are anything but simple though, and thus approximations are almost always a necessity. The topic of this thesis is approximate Bayesian inference and its applications in three intertwined problem domains. Variational Bayesian learning is one type of approximate inference. Its main advantage is its computational efficiency compared to the much applied sampling based methods. Its main disadvantage, on the other hand, is the large amount of analytical work required to derive the necessary components for the algorithm. One part of this thesis reports on an effort to automate variational Bayesian learning of a certain class of models. The second part of the thesis is concerned with heteroscedastic modelling which is synonymous to variance modelling. Heteroscedastic models are particularly suitable for the Bayesian treatment as many of the traditional estimation methods do not produce satisfactory results for them. In the thesis, variance models and algorithms for estimating them are studied in two different contexts: in source separation and in regression. Astronomical applications constitute the third part of the thesis. Two problems are posed. One is concerned with the separation of stellar subpopulation spectra from observed galaxy spectra; the other is concerned with estimating the time-delays in gravitational lensing. Solutions to both of these problems are presented, which heavily rely on the machinery of approximate inference.