Encoding the latent posterior of Bayesian Neural Networks for uncertainty quantification