Hybrid SORN Support Vector Machine Classifier: A Hardware Implementation of the RBF Kernel

In this paper, a new hardware-friendly implementation of a Support Vector Machine (SVM) classifier using the radial basis function (RBF) kernel is presented. It is based on Sets-Of-Real-Numbers (SORN), an interval based number format that allows fast and low complex computing, while providing correct results based on interval arithmetic. For the proposed Hybrid SORN SVM implementation a combination of SORN and fixed point (FxD) representations is used to provide high classification accuracy on low computational complexity. The hardware architecture is implemented on an FPGA board using a 14 bit SORN datatype combined with a 16 bit FxD representation for Euclidean distance calculation, showing an accuracy of 97.71% for the full MNIST dataset, which is only 0.9% below floating point. In addition to RBF kernel, a new SORN-friendly kernel is presented. For this architecture an 18 bit SORN datatype combined with 16 bit FxD is used. Compared to the Hybrid SORN RBF kernel, it consumes less hardware resources including no DSPs, while only increasing the classification error by 0.45%.