The first work, I linked, arguably gets around this issue completely. There we train a neural network to approximate Bayesian prediction directly. It accepts the dataset as input. Thus, there is no optimisation procedure per dataset, just a forward pass. So far, this seems to be pretty stable.