FITC GP approximation with Stan?

[Janne Sinkkonen]

I have logistic spatial data with a linear (non-spatial) covariate, and would like to like to model it with a GP (with unknown scale). The number of data points is around 300–600, and a straightforward logistic GP model with Stan seems slow. (I have tried it, with reparameterizations, jitter, etc.). 

Is there any reason why the FITC approximation would not be feasible with Stan?

The matrix computations seem slightly complicated, and my main worry is that gradients would not be that fast.

Any other recommendations? (I'm not very familiar with GP's.) I mainly care about the mean, so would something simpler be sufficient? Would GPstuff be able to do GP's with linear covariates?

[Aki Vehtari]

On Sunday, April 19, 2015 at 7:00:59 PM UTC+3, Janne Sinkkonen wrote:

I have logistic spatial data with a linear (non-spatial) covariate, and would like to like to model it with a GP (with unknown scale). The number of data points is around 300–600, and a straightforward logistic GP model with Stan seems slow. (I have tried it, with reparameterizations, jitter, etc.). 

Is there any reason why the FITC approximation would not be feasible with Stan?

The matrix computations seem slightly complicated, and my main worry is that gradients would not be that fast.

For that data size the benefit of FITC is likely to be small unless number of inducing points is very small. Currently autodiff is likely to be slow. There is a project for improving GPs in Stan, but it's going to take some time...

 

Any other recommendations? (I'm not very familiar with GP's.) I mainly care about the mean, so would something simpler be sufficient? Would GPstuff be able to do GP's with linear covariates?

GPstuff can handle linear covariates easily. It's likely that you could also get fast good results with EP instead of MCMC for latent values. You can email me if you want more help with GPstuff.

Aki