Novel approximations for inference and learning in nonlinear dynamical systems