Advances in nonlinear model reduction: least-squares Petrov--Galerkin projection and machine-learning error models