B0385
Title: Neural Bayes estimators for irregular spatial data using graph neural networks
Authors: Matthew Sainsbury-Dale - University of Wollongong (Australia)
Jordan Richards - King Abdullah University of Science and Technology (Saudi Arabia)
Andrew Zammit Mangion - University of Wollongong (Australia) [presenting]
Raphael Huser - King Abdullah University of Science and Technology (Saudi Arabia)
Abstract: Neural Bayes estimators are neural networks that approximate Bayes estimators. They are fast, likelihood-free, and amenable to fast bootstrap-based uncertainty quantification. Currently, neural Bayes estimators for spatial models are only available for gridded data. The estimators are also conditional on the sample locations, and need to be re-trained whenever the sample locations change; this renders them impractical in many applications. Graph neural networks are employed to tackle the important problem of spatial-model-parameter estimation from arbitrary sampling locations. The architecture leads to substantial computational benefits since training of the neural Bayes estimator now only needs to be performed once for a given spatial model, and can be used with any number or arrangement of sampling locations. The methodology is illustrated on a range of spatial models, including Gaussian processes and max-stable processes for spatial extremes, which have an intractable likelihood function.
