A0647
Title: Neural methods for likelihood-free inference in spatial and spatiotemporal models
Authors: Matthew Sainsbury-Dale - King Abdullah University of Science and Technology (Saudi Arabia) [presenting]
Andrew Zammit Mangion - University of Wollongong (Australia)
Jordan Richards - School of Mathematics, University of Edinburgh (United Kingdom)
Raphael Huser - King Abdullah University of Science and Technology (Saudi Arabia)
Abstract: Methods for making inferences on parameters in statistical models are often based on the likelihood function. However, for many models, the likelihood function is unavailable or computationally intractable. The use of neural networks is discussed to facilitate fast, likelihood-free inference. These methods are "amortized" in the sense that once the neural network is trained with simulated data, inference from observed data is (typically) orders of magnitude faster than conventional approaches. The methodology is illustrated using spatial Gaussian and max-stable processes, and an application to a data set of global sea-surface temperature is showcased. There, the parameters of a Gaussian process model are estimated in 2161 spatial regions, each containing thousands of irregularly-spaced data points, in just a few minutes with a single graphics processing unit.
