A survey of GPU accelerated SVM

Abstract

Support Vector Machines (SVM) is a set of machine learning algorithms that have been widely used in diverse domains. As the volume of data generated by humans and machines increases year by year, the traditional training algorithms for SVM become infeasible for large scale datasets. Mathematical optimization approaches and computing parallel techniques are two popular strategies to accelerate the training process of SVM. Among those parallel approaches, implementing SVM on Graphics Processing Units (GPUs) has become new research and application interest. General used GPUs have been widely adopted to accelerate a lot of traditional algorithms, including SVM and achieved high performance and speedup. In this work, we survey the mathematical optimization algorithms of SVM training process, as well as GPU accelerated implementations of SVM.