Sparse Online Gaussian Processes
These pages present a method for probabilistic (Bayesian) inference
using
Gaussian Processes (GPs). The
main advantage is the
combination of a non-parametric
and a probabilistic framework: GPs are probabilistic
kernel machines.
The online GP inference provides an approximation to the
posterior process is a GP, meaning that one has the possibility of
assessing posterior uncertainty of predictions. The approximations
are such that they allow a sparse representation of
the posterior and it is applicable to a wide variety of models.
Implementation
The code for the algorithms is provided as a
MATLAB package and is built on the
freely available
NETLAB package
(developed by
Ian Nabney)
which implements general pattern recognition algorithms.
The source code and documentation is available for download in the
Code section of
these pages. The "Hello World" example program for GP inference, which
is the regression using Gaussian noise is presented and explained there.
Additionally to the simple example programs, the package also includes
two demonstration programs with GUI.
-
Regression example
In a first example program (
demogp_reg_gui),
one generates noisy realisations of the
sinc
function. The type and amplitude of the noise and the model
parameters can be set manually.
Using the default settings and generating 40 training points, the
result of learning (before the hyper-parameter adaptation) is
visualised in the image below.
-
Classification example
Similarly to the previous section, the program (
demogp_class_gui)
illustrates the Sparse GP inference for binary classification.
Questions, comments, suggestions: contact Lehel
Csató.
