On approximate Bayesian methods for large-scale linear inverse problems