基于GPU的多类支持向量机改进算法

针对支持向量机算法耗时较长的问题,利用并行计算思想,基于图形处理器对多类支持向量机算法——Crammer-Singer算法进行改进,并利用循环展开、数据暂留、缓存和开放运算语言等技术对算法加以实现.分别在4个数据集上对原算法和改进算法进行对比实验,结果表明,改进算法在性能上获得了较大提升.

Improvement of Multiclass Support Vector Machines Based on Graphic Processor

According to the phenomenon that the support vector machine algorithm takes too much time, the idea of using parallel computation was used to solve this problem. Based on this, an improvement of the classic multiclass support vector machine algorithm first proposed by Crammer and Singer was made, and it was realized by using the great parallel ability of graphic processor. Technology of loop unrolling, data staying, cache and open computing language were used for implementing the improved algorithm. The original algorithm and the improved algorithm were executed on the same four datasets. And the experimental results show that the improved algorithm is much better than the original algorithm in performance of time.