a starting seed for the current iteration. Another idea is to solve a relaxation of RMP and find an effective rounding procedure. These two
approaches may improve the computational performance of GBD.
In this work, we did not explore how one can set λ values to obtain
a good classifier. Perhaps the solution to the basic SVM problem can
provide a starting point for choosing a value for λ. This needs to be
explored as GBD-λ run-time requirements are significantly less than
GBD-M.
One way to think of GBD is as follows: at each iteration it finds a
set of support vectors in the sub problem SUB and then RMP is used
to select a subspace that best fits these support vectors. Some of the
concave minimization ideas can be used here to select the subspace,
given the support vectors