2026/11/18(Wed) 11:00 -12:30 五樓第一會議室 5F, 1st Meeting Room
Title
Accelerating cosmological inference using Gaussian Process Regression
Speaker
艾米克博士 (中央研究院物理研究所) Dr. Ameek Malhotra (Institute of Physics, Academia Sinica)Abstract
The framework of Bayesian inference underpins most statistical analyses for early universe cosmology. However, typical parameter inference or model comparison runs can be slow due to requiring prohibitively many evaluations of expensive likelihoods. Gaussian Process Regression (GPR) combined with an active learning strategy can accelerate this process by learning a cheap surrogate of the likelihood in a small number of evaluations and using the surrogate to perform inference. The practical performance of GPR, however, deteriorates as the dimensionality and complexity of the target likelihood increase. Inspired by techniques used in traditional MCMC routines, we develop two complementary approaches to extend vanilla GPR based emulation to higher dimensions. The first uses numerical marginalisation to integrate out nuisance parameters, thereby reducing the dimensionality of the target. The second uses parameter-space rotations to simplify posterior correlations and improve alignment with the GPR kernel. We demonstrate the effectiveness of these techniques by obtaining accurate parameter posteriors and evidence estimates for the $\Lambda$CDM and its extensions using CMB data in a matter of minutes, a task that would take hours to days with traditional methods.
Language
演講語言 (Language): in English