I've found it fairly easy to get high performance GPRs fitted with gpytorch; kernel design is another problem altogether. I think they take more tinkering than

I've found it fairly easy to get high performance GPRs fitted with gpytorch; kernel design is another problem altogether. I think they take more tinkering than an equivalent neural network by way of the Gaussian process prior. In other words, we want a "smooth" function, but how smooth do we really want it?