Discussion on applications

[jfelip]

http://proceedings.mlr.press/v115/izmailov20a.html
If I understood well, it seems that low dimensional sampling methods have good potential also for BDL using subspace inference. In this case they use HMC but it can be possible that the inference landscape is multimodal or has challenging conditions for the typical MCMC sampling methods, that could open the door to AIS or TP-AIS methods to do the inference of the posterior predictive.
Maybe this can be an interesting contribution to BDL that builds on top of subspace inference?