Bayesian treatment of neural networks