PLSSVM - Parallel Least Squares Support Vector Machine

Alexander Van Craen, Marcel Breyer, Dirk Pflüger

Abstract: h2>Abstract</h2><p>Support Vector Machines are used in supervised learning. For large dense data sets, however, even optimized implementations like LIBSVM or ThunderSVM do not scale well on massively parallel hardware: They are algorithmically based on Sequential Minimal Optimization, and we are not aware of a performance portable implementation supporting CPUs and GPUs from different vendors.</p><p>Our Parallel Least Squares Support Vector Machine (PLSSVM) solves both of these issues. First, PLSSVM resorts to the least squares formulation, and thus to an algorithm that is well-suited for massive parallelism. Second, PLSSVM provides a hardware-independent efficient implementation using OpenMP, CUDA, HIP, OpenCL, and SYCL.</p>