Validation data set for approximate GP using PLL

[rhaghi]

Hi,

I am new to GPyTorch and have a question. I have built an approximate GP model with PLL. It works and has good predictions, but my question is, do I need to have a validation data set? If yes, how can I incorporate the validation process into the training process? I understand PLL is different than normal Log Likelihood. It estimates the distribution of the output, and then in the training process, using KL divergence as a loss function at the inducing points, it tries to optimize the fit. So, to my mind, a validation data set, would not help in this scenario as it is a smaller dataset, and the distribution of this smaller data will not match the bigger dataset that PLL wants to estimate. However, I can't find any reference in the literature for my argument. What do you think? Does it make sense?