The Development of FPGA Implementation on Convolutional Neural Network

In this article, we discuss the multifaceted aspects of implementing Convolutional Neural Networks (CNNs) on Field-Programmable Gate Arrays (FPGAs). In view of the rising demands for realtime processing and low-latency machine learning applications, the deployment of CNNs on FPGA presents a promising solution. We scrutinize several existing implementations, focusing on three key components: architectural optimization, memory management techniques, and the computation enhancement technniques. The advantages and limitations of each method are discussed in detail. Furthermore, we explore the unique opportunity provided by FPGAs to locally conduct both feedforward and training processes, promising notable enhancements in data security and privacy. A flexible and robust general structure for FPGA implementation on any CNN is proposed, providing a foundation for future research and development. This article serves as a comprehensive resource, aiming to deepen the understanding of FPGA-based CNN implementations and spur further explorations in this area.