Hi Pyro Community,
Is it possible to assign spike and slab prior to the GP’s lengthscale parameters in Pyro? I am trying to use spike and slab prior to do variable selection in GP, but I am not sure if the GP API supports it. It would be great if any of you have experience or suggestions about using spike and slab prior in Pyro. Thanks a lot!
Can you be more specific? AFAIK spike-and-slab is usually used to learn which predictors to include in a discriminative model (usually seen in the context of some type of GLM). So your question about applying them to length scale parameters is interesting to me. Do you mean something like the following: I have a multi-input GP and would like to apply spike-and-slab methodology to variable selection in the GP? Presumably then you want to apply the “slab” part to the length scale parameters…? If this is something like what you’re thinking about, we can figure it out. FWIW it is obvious that you can do this if you drop the GP API – as your “GP” just use a multivariate normal the dimension of which is determined by the “spike” part.