Appendix A Derivative of ELBO) between the variational approximate posterior q(f,f) and the true posterior p(f,f

We first derive the (6). In this section, we provide implementation details regarding the proposed method. All experiments are conducted on a server with 8 Nvidia RTX 3090 graphic cards, and each has 24GB memory. Every experiment we report can be trained and tested on a single card. The machine is also equipped with 512GB memory and two AMD EPYC 7502 CPUs.