search

UMDataScienceLab / SGD-in-Gaussain-processes

We show that SGD can indeed be used to infer Gaussian processes. This in turn allows GPs to scale far beyond what was thought possible.
3 stars 1 forks source link

Restricting learned noise #2

Closed shivam15s closed 1 year ago

shivam15s commented 2 years ago

The paper mentions "We manually inject observational noise into simulated datasets. For the query dataset, we constrain the learned noise to be at least 0.1 to regularize the ill-conditioned kernel matrix". Can you clarify how much observational noise do you add to the simulated datasets? Also, could you clarify whether the learned noise is constrained only for the query dataset and not for other simulated/real datasets?

UMDataScienceLab commented 2 years ago

In terms of the injected noise, the standard deviation of the noise is 5 for Levy, 4 for Griewank and 5 for Borehole. In terms of constraining learned noise variance, the default is 1e-4 for all datasets other than Query.