Using neural networks to directly learn and approximate the posterior distribution of model parameters.