A hierarchical Bayesian learning scheme for autoregressive neural networks: application to the CATS benchmark