The fortnightly BRAG meeting will be held this Thursday, May 1st , at 1pm via Zoom/GP-Y801. This week we will have a presentation by Adam Bretherton.
Zoom link: https://qut.zoom.us/j/87984815350?pwd=mTrT5HlXwfOCWMZDmi4tCva8aUPXM2.1
Meeting ID: 879 8481 5350
Passcode: 606649
Adam’s Talk:
Title: Flexible transformations for Bayesian Score Calibration
Abstract: To represent real-world systems more accurately we often require complex mathematical models. Unfortunately, such mathematical models often possess likelihood functions that are either intractable or computationally prohibitive. As such, approximate models or simulation-based inference (SBI) techniques are often required. Approximate models use a simpler model (surrogate) with a tractable likelihood that provides only macroscopic information. Alternatively, SBI techniques avoid explicitly evaluating the likelihood by repeatedly comparing the observed data to costly simulations from the data-generating process (DGP). Therefore, we might turn to recalibration methods that attempt to correct the approximate model towards the DGP and require only a small number of model simulations. One such method is Bayesian score calibration which learns a location-scale transformation for the approximate posterior. This transformation applies a static correction to the first two moments across the parameter space. Unfortunately, the misspecification between the surrogate and DGP might change across the parameter space. Therefore we propose several transformations that can handle complex differences between the surrogate and DGP. Further, our extensions to Bayesian score calibration show better performance across several examples.