Study On General Purpose GPU Computing In Classification Algorithms

Driven by the rapidly growing demand for 3D rendering and graphics processing, theGPU (Graphics Processing Unit) has developed into a kind of micro-processor with tremen-dous computational horsepower and highly thread parallelism. GPGPU (General PurposeGPU) is becoming a research focus. This paper concentrates on applying GPU to classifica-tion algorithms. Based on the study of the GPU architecture, the paper presents principlesand methods in designing GPU algorithms using CUDA (Computing Unified Device Archi-tecture). Some typical algorithms are also used to verify these principles and methods. Themain research work of this paper is as follows.Firstly, the paper discusses the feasibility of GPU in classification algorithms. Theanalysis and appraisal show that better results could be achieved on the condition that somedesign constraints using CUDA platform is met.Secondly, the paper proposes GSNN algorithm, which is a GPU-based segmentationnearest neighbor classification method. The algorithm uses a segmentation strategy in dis-tances calculation and an execution assessing method in nearest neighbor selection.Thirdly, a GPU based massively data parallel C-SVM classification (GMP-CSVC) al-gorithm is presented to reduce the training time of SVM.Finally, based on the improvement algorithm ofν-SVM, GMP-nuSVC is proposed tosolve the performance issue in cross-validation and parameter selection of SVM. A cachescheduling scheme and a tiling method of calculation are used in the algorithm.The paper successfully applies GPU to the field of classification algorithms by extend-ing the border of calculation and reducing the training time. It has certain practical signifi-cance in the application of classification methods and the research of GPGPU algorithms.