55 min
Seminar, Hosted by One World Approximate Bayesian Inference (OWABI) Seminar
Multilevel Neural Simulation-Based Inference
University College London
Recording
Abstract
Neural simulation-based inference (SBI) is a popular family of methods for Bayesian inference when models are only available as simulators, widely used in the sciences and engineering where writing down a likelihood is far harder than building a simulator. Its performance suffers, however, when simulators are computationally expensive and only a limited number of simulations can be run.
This talk presents a new approach to neural SBI that uses multilevel Monte Carlo techniques in settings where several simulators of varying cost and fidelity are available. Theoretical analysis and extensive experiments show that the method can substantially improve the accuracy of SBI for a fixed computational budget (Hikida, Bharti, Jeffrey and Briol, arXiv:2506.06087).