search

gpstuff-dev / gpstuff

GPstuff - Gaussian process models for Bayesian analysis
http://research.cs.aalto.fi/pml/software/gpstuff/
GNU General Public License v3.0
169 stars 61 forks source link

EP and Laplace latent models with DTC-SOR-VAR sparse approximation #53

Open andreapatane opened 3 years ago

andreapatane commented 3 years ago

Hi, I am not sure whether the software is still mantained, but anyway.

I'm having troubles using DTC, SOR and VAR sparse approxiamtions with the EP latent model for classificaiton. I believe that the problem comes from a missing sign in line 767 of gpep_pred.m Currently it says:

Varft = sum(B2'.(B(repmat(La,1,m).\B')B2)',2) + sum((K_nu(K_uu(K_fu'*L))).^2, 2);

But this gives negative variance values. I think the correct formula might be:

Varft = sum(B2'.(B(repmat(La,1,m).\B')B2)',2) - sum((K_nu(K_uu(K_fu'*L))).^2, 2);

But I don't have a great knowledge of sparse GP, so I might be wrong and the problem might be something else.

Also the Laplace method doesn't work with any sparse methods, and I get some matrix dimensionality problem, but I haven't investigated much.

Thank you!

jpvanhat commented 3 years ago

Hi,

I would like to know a bit more what you want to do. Could you send me a short example of where things break down. EP and Laplace should work with these approximations.

-Jarno