Seminario di John Veitch

Host: D. Gerosa

 

Bayesian modelling using normalising flows Bayesian inference has long been used for gravitational wave analyses, from estimating parameters of individual sources to inferring hierarchical models of the source population. Bayesian models must be explicitly written down in terms of their parameters, be they physical or phenomenological, which are typically estimated with a stochastic sampler. This procedure has served us well, but has its limitations: the functional form of the model must be specified in advance, and we cannot directly evaluate the posterior density without reanalysing all data. Normalising flows are a machine learning technique to describe an arbitrary density function. I will explain how they can be used to address the issues with traditional models while retaining their advantages. I give some examples from recent work on source population and glitch modelling, and show how flows can enable true Bayesian updates.