There are several methods closely related to Gaussian Processes
that are relevant to this work. Wahba [31, 29] makes
a connection between Gaussian processes and reproducing
kernel Hilbert spaces (RKHS), showing that the solution
to the posterior Bayesian estimate of the Gaussian process
(as in Equation 4) is also the solution to a spline smoothing
problem posed as a variational minimization problem
in an RKHS. The smoothing parameter is optimized using
cross validation methods, whereas in the case of Gaussian
process priors, we use a MAP estimate for the hyperparameters.
In related works (e.g. [30]), the relevance of the
different components of the function is estimated from the
learned smoothing parameters. In Gaussian process methods
there is a similar notion, judging the relevance of different
input dimensions by their estimated lengthscales in