Rethinking Approximate Gaussian Inference in Classification