Generative modelling for simulation-based inference

Simulation-based inference (SBI), traditionally also known as likelihood-free inference, refers to a class of statistical problems where generating data samples via simulation is possible, but the likelihood function is computationally intractable. Such problems appear in a diversity of real-world application domains such as engineering, economics, and many scientific fields. Our research aims to enhance SBI by leveraging breakthroughs in generative modelling.

There have been significant advancements in the field of generative modelling in recent years, offering new tools for probabilistic modelling. From a methodological angle, improvements in probabilistic modelling have much potential for cross-pollination with SBI methods, which have increasingly incorporated machine-learning techniques. However, many challenges remain, especially concerning sample-efficient inference procedures and the robustness of inference.

Specifically, our research aims to incorporate novel generative modelling techniques, such as score matching for gradient estimation, to improve the inference procedure of SBI. We hope to be able to leverage modern machine learning methods to address the various challenges facing SBI methods and to broadly improve and expand the statistical toolkit for SBI.