About Simulation-Based Inference

The problem of inference is ubiquitous in modern science. In the past decades, many statistical tools have seen their use consolidated across different scientific fields, either through frequentist or Bayesian approaches.

In parallel, the design of more powerful and more efficient computing hardware has allowed for the design of high-fidelity simulations of physical systems with increasing complexity, allowing for the generation of synthetic data from them. However, performing inference from these simulators still remains challenging. In these contexts, the likelihood function is not explicitly calculated, and is instead implicitly defined by the data-generating process implemented by the simulator. The problem of performing inference with these systems has thus been named likelihood-free inference or simulation-based inference (hereafter SBI).

In the past few years, the development of more sophisticated Machine Learning techniques, in particular deep neural networks, and the production of specialized hardware for training has given new momentum to the field of SBI. For a high-level overview of the impact of these trends on the emergence of new methods, we refer the reader to this review paper by (Cranmer et al, 2019).

For more information on applications of SBI in scientific problems, see our resources section.

References

[1]: Cranmer, Kyle, Johann Brehmer, and Gilles Louppe. "The frontier of simulation-based inference." Proceedings of the National Academy of Sciences 117.48 (2020): 30055-30062.