A framework for fast multidimensional GP analysis through additive modeling and SSM analysis.
To read about some preliminary experiments using this code, see:
by Gilboa, Elad, Yunus Saatçi, and John P. Cunningham. "Scaling Multidimensional Inference for Structured Gaussian Processes." to appear TPAMI (2013).
Feel free to email with any questions: [Elad Gilboa] ([email protected]) [Yunus Saatchi] ([email protected])
You'll need Matlab and GPML. For comparisons you might need to install GPstuff-3.3, IVM, and libsvm. For compilation of the mex files you will also need the boost_1_45_0.
Before running additiveGP you must call config.m to setup the
paths. You need to change config.m for your lib locations.
To check whether the framework runs, go to the source directory and run 'reg_runtime_N_comparison' for regression and 'run_breast' for classification.
There are some example experiment scripts examples/.
for regression:
numSubset = 1000; %subset of data to use for MCMC inference dproj = D; %number of projection dimensions
numPseudo=500; %number of pseudo inputs for SPGP
numMCMC = 10; %number of full MCMC iterations
rand_init = false; %initialize proj pursuit weight randomly or with linear model
for classification:
addLA.runMCMC;% whether to run the approximation with MCMC or Laplace approximation
addLA.ells; %length scale hyperparameter for gpml;
addLA.sigfs; %variance hyperparameter for gpml;
addLA.numNewton; %number of Newton iterations;
addLA.numGS;% number of Gibbs sampling for posterior calculation
If you have any questions about getting this running on your machine or cluster, please let us know.