FPGA implementation of a feedforward neural network-based classifier using the xQuant technique

摘要

For decades, feedforward neural network (FNN) based classifiers have been extensively used in many classification problems such as image and speech recognition. The inherent parallel nature of the FNN classifier makes it a good candidate for hardware implementations in order to obtain a performance speed up, since most of its computations are matrix-vector multiplications, where the input images arranged as vectors are multiplied by a matrix of parameters learned for a specific set of images. Nevertheless, when the number of parameter of a FNN classifier is large, embedding it in FPGA becomes a challenging task. This work presents the implementation of a quantisation strategy called xQuant in hardware. xQuant was developed to reduce the FNN implementation cost by approximating floating-point parameters by its nearest power of two and replacing floating-point multiplications by bit shift operations. Using the FNN classifier learning algorithm LAST (Learning Algorithm for Soft-Thresholding), a classifier for a texture classification problem was trained and used as study case. Both the floating-point and the xQuant architectures were implemented on a Xilinx Zynq-7020 SoC, and results show a three fold increase in the performance and a large reduction in FPGA resources: 4x reduction of LUTs; 7.4x reduction in FF; 10x reduction in RAM blocks and the elimination of DSP blocks.