Speeding up LP-SVM Using CUDA Platform

In this paper, CUDA platform is used to speed up the LP-SVM (linear programming - support vector machine), so that the training algorithm can handle large-scale data on PC (personal computer). Considering the limited memory of PC, we modified the decomposition algorithm of SVM to solve the LP-SVM. Each time the new decomposition algorithm only tackle a small-scale linear programming problem and avoids loading all the training data into memory. The most time-consuming parts of linear programming are migrated to CUDA platform, which significantly improved calculation efficiency. Experimental results show that the algorithm efficiency has increased by 10 to 35 times.