Research on Image Classification Algorithm Based on Dynamic Decay EMA
YANG Jing-dong,ZHU Jin-tu,SUN Xin-bo,YANG Wen-hao
(Autonomous Robotics Laboratory,School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)
Abstract：As the traditional exponential moving average(EMA)algorithm cannot optimize the network parameters during training later period continuously,the overfitting often happens for the deep convolutional neural networks.In this paper we propose a dynamic attenuation index moving average algorithm(T-ADEMA),which uses the Tanh function with variable coefficient as the attenuation function,and adjusts the optimization parameters according to training times dynamically,Further to reduce the noise of datasets,and improve the generalization performance.Experiments show that the deep network based on T-ADEMA algorithm achieves higher accuracy than the traditional EMA algorithm on the three data sets of MNIST,CIFAR_10,CIFAR_100(20 and 100 categories).
杨晶东,朱锦图,孙新博,杨文皓. 基于动态衰减EMA的图像分类算法研究[J]. 小型微型计算机系统, 2020, 41(7): 1524-1529.
YANG Jing-dong,ZHU Jin-tu,SUN Xin-bo,YANG Wen-hao. Research on Image Classification Algorithm Based on Dynamic Decay EMA. Journal of Chinese Computer Systems, 2020, 41(7): 1524-1529.
2020, Vol. 41 
