% The following information is kindly provided by Tomer Hertz. 

In order to run this code in Kernel mode you should set the following
parameter within the BSParams struct (which contains all of the
relevant parameters for the code to run, most of which have default
values):
BSParams.MarginType = 1; %when set to 0 this runs DistBoost, when set
to 1, runs KernelBoost without the label dissolving mechanism (Which
also helps, as shown in the paper).

Other relevant parameters which you should take notice of are:

BSParams.T - the number of boosting rounds.

BSParams.WeakParams.M - the number of models in the weak learner.
This should be set to be no less than the number of labels/classes in
the problem you are working on. The weak learner has an automated
mechanism which can remove models that are unnecassary (i.e. who do
not "explain" any of the data points in the training set), so you can
choose more models than needed, and the weak learner will discard
some of these if they are not needed. Best performance is usually
obtained when selecting the number of models M to be a little higher
than the number of classes in the dataset.

BSParams.WeakParams.single_cov_mat_flag - determines the number of
covariance matrices used by the weak learner. The default is set to
1, which usually works best, since the number of parameters to be
estimated is significantly lower and is not dependent on the number
of classes in the dataset. Note however, that on some datasets in
which the distribution of the classes is very different from one
another, this flag can be set off.