Bit-serial convolution with prediction threshold for convolutional neural networks: Electrical Engineering Subject Index: EL7 Signal Processing. (3rd April 2022)

• Record Type: Journal Article
• Authors: Hsiao, Jen-Hao; Chin, Wen-Long; Wu, Yu-Feng; Chang, Deng-Kai
• Is Part Of: Journal of the Chinese Institute of Engineers. Volume 45:Number 3(2022)
• Page Start: 266
• Page End: 272
• Publication Date: 2022-04-03
• Subjects: Convolutional neural network -- hardware accelerator -- serial input; Technology -- Periodicals; Engineering -- Periodicals; 620.005
• DOI: 10.1080/02533839.2022.2034050
• Languages: English
• ISSNs: 0253-3839

Abstract

ABSTRACT: To reduce the implementation complexity and power consumption of the convolution operation in a convolutional neural network (CNN), this work proposes a new convolution method using the serial input and prediction threshold. To confirm the benefits of the proposed method, we use the original parameters, i.e. kernel weights and biases, of the popular AlexNet to verify the proposed algorithm and then implement its digital circuit. According to implementation data and comparison with traditional convolution using bit-parallel input, the implementation gain in terms of the throughput/power/area of the serial convolution is 7.57 times that of parallel convolution.