This should follow closely the implementation of Approx. Power EP for GPSSMs, and Deep GPs for regression with alpha=1 with no hidden variables.