Speaker
Description
Suppose you have some observational data which is well modeled by a complex simulator. The inverse problem is: "Which parameters could be input into the simulator to reproduce the observational data?" Simulation-based Inference estimates the probabilistic solution to the inverse problem by creating a surrogate model for the posterior distribution.
We created a generalized method to estimate the posterior distribution given flexibly drawn training (i.e. simulation) data. We take advantage of extremely expressive estimators that take the "best-of-both-worlds" compared to other simulation-based inference methods. Our method is extremely data efficient by using a favorable estimator design in the "bias-variance tradeoff" sense. We present the method itself and results on various benchmarks.
