Design and Implementation of Searching SVM Optimal Training Parameter Set Based on Shared Dot Product Matrix

Abstract

The Optimal Training Parameter Combination (OPTC) is the core of the support vector machines (SVM) to construct application model. However, the calculated amount of searching OTPC of SVM is extremely huge, which is time-consuming during the process of implementation by the software. To solve this issue, we propose a Shared Dot Product Matrix (SDPM) algorithm. The algorithm computes the dot product of all training data sets and stores them simultaneously, which achieves an ultra-fast processing speed. Meanwhile, the hardware/software co-design architecture for searching OTPC of SVM is proposed to corporate the data processing. The implementation and test results have shown that, the software and hardware collaboration system proposed in this paper has the performance that is 30 times faster in searching speed than software-based LIBSVM.