A0783
Title: Normalizing flows for posterior estimation under intractable likelihoods, with applications in astrostatistics
Authors: Roxana Darvishi - Simon Fraser University (Canada) [presenting]
David Stenning - Simon Fraser University (Canada)
Owen Ward - Simon Fraser University (Canada)
Abstract: Normalizing flows are a class of models that allow for flexible density estimation and efficient sampling of complex distributions. These models use neural networks to learn a transformation that maps a simple random variable to the target of interest. In the context of Bayesian hierarchical models, normalizing flows are particularly useful when the likelihood function is unavailable or computationally expensive to evaluate, limiting the applicability of conventional inference techniques. A two-stage algorithm using normalizing flows is proposed to enable density evaluation and sample generation when the target distribution can be factorized into components that are either available in closed form or accessible through sampling. The method is applied to a challenging astrophysics analysis for which case-by-case posterior samples for several physical parameters of thick-disk white dwarf stars are available, but the explicit mathematical forms of the underlying densities are unknown. The goal is to approximate the full joint posterior and enable both sampling and density evaluation simultaneously. The effectiveness of the algorithm is evaluated by comparing it to existing methods, highlighting its potential advantages for dealing with intractable likelihoods. Broader applications are also discussed, including joint density estimation and reducing the computational cost of sample generation.
