Articles • neuralsbi

Introduction to neural simulation-based inference in R: The basics of amortized, likelihood-free Bayesian inference: prior, simulator, observation, and how neuralsbi turns them into a posterior, illustrated with a simulator whose likelihood has no closed form.
Getting started with neuralsbi: Train a neural posterior estimator from a prior and a simulator to perform amortized, likelihood-free Bayesian inference in R, then check that the fit is trustworthy.
Choosing a density estimator: Compare the conditional density estimators behind neural posterior estimation in R (mixture density network, masked autoregressive flow, and neural spline flow) and pick one for your simulator.
Checking the posterior: Validate a neural posterior in R with simulation-based calibration (SBC), expected coverage, TARP, and posterior predictive checks before you trust its credible intervals.
neuralsbi and pomp: two routes to an SIR posterior: A side-by-side comparison on a stochastic SIR epidemic: infer the transmission and recovery rates with pomp’s particle-filter MCMC and with neuralsbi’s neural posterior estimation, and check whether the two posteriors agree.
Amortized R(t): behavioral SIR models across all US states: Fit three competing epidemic models to daily reported cases in every US state, summarize the 120-day case curve with a learned embedding network, and ask which behavioral mechanism explains the shape of the first SARS-CoV-2 wave.