Reconstruction and prediction of nonlinear dynamical systems: a hierarchical Bayes approach with neural nets