Reduction of memory footprint and computation time for embedded Support Vector Machine (SVM) by kernel expansion and consolidation

Support Vector Machines (SVM) are a family of algorithms widely used in classification and regression tasks. When faced with large and structurally complex data sets, the classification or prediction by SVM can become memory and time intensive, and especially so in non-sparse variants of the SVM such as the least-squares SVM (LS-SVM). This is of particular importance for implementing classifiers in embedded systems where memory and computation capabilities are limited. In this paper, decomposition methods for SVM classification functions are developed and discussed, using polynomial approximation methods. The SVM decision function is expanded into a polynomial form and consolidated into classification function with a significantly lower memory footprint and computational cost. The amount of reduction is analyzed for polynomial kernels, and in three demonstrated example systems, the classifier is made two orders of magnitude faster, with a memory requirement that is two orders of magnitude smaller. The methods are tested on standard classification data sets, and guidelines for use of the methods are provided.