I'm trying to make up a taxonomy of UQ methods for deep learning models, if possible (this paper provides a nice overview imo, albeit in a specific field).
Currently there's cluster of UQ approaches I'm naming 'sampling' which all run multiple forward passes during inference (Bayesian DL, TTA, MC dropout, ensembling). Afaik, MC drop-out is considered (at least by some) to be an approximation of Bayesian DL, and also somehwat related to model ensembling. What I wanna ask is, how does Test-time augmentation (TTA) fit into this picture? Bayesian DL tries to get a distribution of model parameters (and MC dropout tries to approximate said distribution) and TTA tries to sample a distribution of input values (with the same y_true). Can i (loosely) claim TTA is also a Bayesian DL approach?
