We invite MIDAS members to explore the newly-released neuralsbi R package. neuralsbi provides neural simulation-based inference (SBI) methods for R-native researchers doing inference on mechanistic models.
We are interested in seeing folks testing neural SBI methods alongside their favorite inference approaches (iterative filtering, ensemble kalman filters, ABC-SMC and the like). One differentiating feature of neural SBI methods is that they support “amortized” inference: e.g., pay an upfront “cost” to train your estimator, then get a posterior back for all the cities/counties/states you are fitting the model to instantly (see getting started vignette for an SIR example). Because amortized inference is fast, you can then do a much better job at following a proper Bayesian workflow, like running simulation-based calibration checks.
This is an open-source project and we are interested in finding early adopters and collaborators. If that’s you, reach out to Pedro Nascimento de Lima at plima@rand.org.
