Scalable posterior inference for the BDWI model

Investigate whether divide-and-conquer methods can be adapted to reduce the computational cost of auxiliary simulations and improve the scalability of posterior inference for the Bayesian distance-weighted Ising model, particularly for large neuroimaging datasets.

Background

The Bayesian distance-weighted Ising model uses auxiliary simulations within each MCMC iteration to address its intractable, parameter-dependent normalising constant. Although the proposed approach performs well for the datasets considered, these auxiliary simulations may become computationally expensive as the dataset size increases. The authors therefore identify the adaptation of divide-and-conquer methods as an unresolved direction for improving the scalability of posterior inference.