Research on Machine Learning Optimization Algorithm of CNN for FPGA Architecture

Abstract

In order to meet the application requirements of deep learning detection and recognition algorithm for Field Programmable Gate Array (FPGA) computing structure, this paper proposes a machine learning optimization algorithm of convolutional neural network (CNN) based on pruning and Int8 quantization. By analyzing the machine learning intelligent recognition network, the Yolo v3 target detection network is selected to verify the optimization algorithm. Aiming at the Yolo v3 network with balanced speed and precision, the optimization algorithm is studied in detail. Based on Amazon Cloud FPGA EC2 instance platform and ZCU104 FPGA hardware platform, machine learning optimization algorithms are used to accelerate the application of Yolo v3 network detection and recognition. Satisfactory results are obtained on both FPGA computing platforms. The experimental results show that the machine learning optimization technology can improve the inference speed of neural network based on both FPGA platforms. The network structure optimized by pruning and Int8 quantization algorithms can achieve high acceleration speed, with very limited accuracy loss.