Trusting posteriors learned from simulations
A neural network trained on simulations can return a posterior in seconds. Start with what a posterior and its 90% region promise, learn the rank and coverage tests that check the promise, and then see why a network can pass every test and still be wrong about real data.
L1 · Explainer
How can you tell whether a neural network's posterior deserves trust?
Test it on simulated data where the answer is known, compare it with an exact calculation, and check that the real data look like the simulations.
L1 · Chapter
What a posterior and a 90% credible region promise
Build a posterior from a prior and a likelihood, find its 90% region on a grid, and see why the region catches the truth nine times in ten.
L2 · Chapter
Checking posterior calibration with ranks and coverage
Simulation-based calibration ranks the true value among posterior draws. Flat ranks mean calibrated; humps, cups and slopes name the failure.
L3 · Chapter
Finding a wrong simulator with importance weights and residuals
Importance weights test the network against the assumed likelihood; a residual test checks the likelihood against the data. A hidden second mode can pass both.
L3 · Method
Recompute the ringdown posterior calibration
Run simulation-based calibration on a conjugate Gaussian model, then on a frozen neural posterior for a synthetic ringdown, with saved inputs and failure cases.
L4 · Paper guide
Reading the DINGO evidence: P–P plots, divergences and sample efficiency
A figure-by-figure guide to the 2021 DINGO letter and the 2023 importance-sampling letter, with what each test shows and how out-of-range data are flagged.