Imposing functional priors on Bayesian neural networks