Proposal Description

Amortized simulation-based inference has emerged as an effective framework for performing statistical inference in complex models whose likelihood functions are unavailable or prohibitively expensive to evaluate. Rather than relying on explicit likelihood calculations, these methods learn mappings from simulated data to parameters or posterior quantities of interest. After an initial set-up cost involving simulation from the statistical model, they deliver inference almost instantaneously, making it feasible to analyse large data sets or embed inference within larger workflows that would otherwise be computationally out of reach. In doing so, amortized approaches bring together ideas from classical statistics and modern deep-learning methodology, providing flexible tools that apply across a wide range of modelling settings.

This session aims to highlight recent developments in amortized simulation-based methods, with a focus on understanding the settings in which they perform well, diagnosing and improving their reliability, and identifying the types of scientific and applied problems that stand to benefit most from their use. The goal is to foster discussion about both methodological foundations and practical considerations, and to clarify the opportunities these techniques offer for fast, scalable, and well-supported inference in demanding applications.

The session will involve a diverse team of participants spanning a range of academic careers. Lídia André will discuss amortized neural methods for both model selection and parameter inference in flexible multivariate extremes models where likelihoods are unavailable or expensive to compute. Matthew Sainsbury-Dale will discuss and contrast two strategies for applying amortized neural methods in the ubiquitous case where data contain missing values: a masking approach and an expectation-maximization (EM) approach. Paul Bürkner will discuss related topics to be determined later.

The session will be chaired by Raphaël Huser, with discussion provided by Andrew Zammit-Mangion. Both have extensive experience with amortized simulation-based inference and have contributed to its methodological development in recent work.