I agree choosing priors is hard, but choosing ReLU versus LeakyReLU versus sigmoid seems like a problem with using neural nets in general, not Bayesian neural nets in particular. Am I misunderstanding?