Review on FPGA-based accelerators in convolutional neural network

Abstract

In current research, an important means to realize artificial intelligence is artificial neural network. Many tedious problemscan be solved by artificial neural network, such as image classification, speech recognition, natural language processing. Among neural networks, although convolutional neural network has significantly better performance in the field of image recognition, it requires many parameters and a large amount of computation, and the number of network layers is gradually increasing with the progress of the algorithm, which leads to model sizes are getting larger and more difficult to handle, which has more stringent requirements on various aspects of hardware capabilities, such as computing power, data storage and memory bandwidth. Therefore, the research of acceleration is particularly important, especially in the field of hardware acceleration. And as an emerging hardware platform, field programmable gate array has the characteristics of rapid customization and programmability, which can greatly improve the efficiency of accelerators design, implementation, and verification. As a result, it can be widely used in the hardware-accelerated design of neural networks. Although field programmable gate array itself has some defects, many research results have proved its huge advantages and development potential. This paper mainly summarizes the acceleration of convolutional neural network based on field programmable gate array platform, and discusses its characteristics and application space by comparing with other platforms, showing its high applicability to convolutional neural network, and finally looking forward to the progress of related research work.