On the effectiveness of neural priors in modeling dynamical systems