Dual Parameterization of Sparse Variational Gaussian Processes A Tighter Bound for the M-step We here study the role of parameterizations ξ

For a matched optimal E-step, i.e. We here detail the computations required to perform inference and learning using the dual parame-terization. To perform inference, the variational expectations need to be evaluated. Eq. (8) needs to be evaluated which requires the computation of a KL divergence. We used a softmax likelihood with 10 latent GPs, one for each digit.