solver.  This makes some things, such as the structural_object_detection_trainer,
significantly faster.

non-negative weight learning support.

suggested by Platt's paper.

sparse vectors to avoid unnecessary RAM usage in this case.