Evaluate gradient for kernel parameters (except bias) for the Sparse GP
covf = ogpcovgrad(net,xTrain);
g = ogpcovgrad(net)
takes a Sparse OGP data structure net
and
evaluates the gradient g
of the negative log-likelihood with respect
to the hyperparameters of the model. The output g
is a matrix with
columns of length #BV^2
(the kernel matrix is put into vector format). Each
column in g
corresponds to a hyper-parameter and the order is provided
by ogppak
.
g = ogpcovgrad(net,xTrain)
additionally to the data structure
net
, the procedure takes the elements of the training set xTrain
and returns on the columns the derivatives of the kernel function. The
result is thus a matrix with (nBV x nTr
) lines and the number of
columns the number of hyperparameters.
If the kernel is specified by the user - i.e. kernel type is 'USER'
-
then the last group of parameters is given by the user and there must exist
a field 'fngrad'
in the structure 'net.kpar'
. For details on
specifying a different kernel see the documentation for function ogpcovarf
.
ogp
, ogppak
, ogpevidgrad
, ogptrain
Copyright (c) Lehel Csató (2001-2004)