2025 workshop program
Robust Simulation-Based Inference: Bridging the Gap Between Simulation and Observation.
The work
Abstract
Recent advances in neural density estimation have enabled powerful simulation-based inference (SBI) methods that can flexibly approximate Bayesian inference for intractable stochastic models. Although these methods have demonstrated accurate posterior estimation when the simulator accurately represents the underlying data generative process (DGP), they have been shown to perform poorly in the presence of model misspecification. In this talk, we address this challenge by discussing error modelling for robust SBI, which introduces a model to account for the simulation gap between the true and simulated DGP. We give a background to previous developments in error modelling for SBI and propose neural latent prior estimation (NLPE). NLPE is a framework for SBI which assumes that the simulator produces a latent representation of the data-generating process, and the gap between the simulator and observations is given by a family of distributions whose parameters are to be inferred. We demonstrate that NLPE can produce robust posteriors on numerical tasks with flexible error models and provide examples of how to use NLPE in astronomy.