Some Improvements to a Parallel Decomposition Technique for Training Support Vector Machines