Assessment of amortized fully Bayesian neural methods (BayesFlow, JANA) for extreme-value applications

Investigate the applicability and performance of amortized fully Bayesian neural inference methods, specifically BayesFlow and JANA, for parameter inference and uncertainty quantification in extreme-value applications involving spatial or spatio-temporal extremes.

Background

The authors point to amortized fully Bayesian neural methods as promising tools to address challenges in flexible spatial tail dependence modeling, including high-dimensional settings and computationally intensive likelihoods.

However, their suitability and reliability in extreme-value contexts have not yet been established. A careful investigation is needed to assess accuracy, robustness, and practical utility relative to traditional likelihood-based approaches, especially for extrapolation to joint tail regions.