Posterior Refinement Improves Sample Efficiency in Bayesian Neural Networks Appendix A Derivation of the Multi-Class Probit Approximation