OptQuant: Distributed training of neural networks with optimized quantization mechanisms

doi:10.1016/j.neucom.2019.02.049

CORC > 数学与系统科学研究院 > 中国科学院数学与系统科学研究院 > 应用数学研究所

	OptQuant: Distributed training of neural networks with optimized quantization mechanisms
	He, Li 2,3; Zheng, Shuxin 4; Chen, Wei 1; Ma, Zhi-Ming2,3 ; Liu, Tie-Yan 1
刊名	NEUROCOMPUTING
	2019-05-07
卷号	340 页码:233-244
关键词	Distributed machine learning Deep learning Stochastic gradient descent Gradient quantization
ISSN号	0925-2312
DOI	10.1016/j.neucom.2019.02.049
英文摘要	Nowadays, it has been a common practice to speed up the training of deep neural networks by utilizing multiple computational nodes. It is non-trivial to achieve desirable speed-up due to the potentially large communication overhead in distributed training. To reduce the communication cost, several unbiased random quantization mechanisms were proposed, in which the local workers, i.e., computational nodes, quantize their local gradients before communications with other workers. Most previous quantization mechanisms are static, i.e., the gradients are quantized in the same way during the training process. However, for different neural network models, the distributions of gradients might be very different even after normalization. To minimize the quantization loss better, we design the parameterized unbiased quantization mechanisms and dynamically optimize the quantization mechanism during the training process, using the aggregated information of the gradients. We call the distributed deep learning algorithms with our new quantization method (unbiased) OptQuant algorithms. Theoretically, we show that the unbiased OptQuant algorithms converge faster than the static unbiased quantization. In addition, if we trade-off the bias and the variance in the quantization, the algorithm converges faster. Motivated by this theoretical result, we further design the parameterized biased quantization mechanisms and the biased OptQuant algorithms. We evaluate our algorithms for different deep neural networks with benchmark datasets. Experimental results indicate that the OptQuant algorithms train the neural network models faster than previous quantization algorithms and much faster than the float version. (C) 2019 Elsevier B.V. All rights reserved.
资助项目	National Center for Mathematics and Interdisciplinary Sciences of Chinese Academy of Sciences
WOS研究方向	Computer Science
语种	英语
出版者	ELSEVIER SCIENCE BV
WOS记录号	WOS:000461360000020
内容类型	期刊论文
源URL	[http://ir.amss.ac.cn/handle/2S8OKBNM/33374]
专题	应用数学研究所
通讯作者	He, Li
作者单位	1.Microsoft Res, 5 Dan Ling St, Beijing, Haidian Distric, Peoples R China 2.Univ Chinese Acad Sci, 19 A Yuquan Rd, Beijing, Shijingshan Dis, Peoples R China 3.Chinese Acad Sci, Acad Math & Syst Sci, 55 Zhongguancun East Rd, Beijing, Haidian Distric, Peoples R China 4.Univ Sci & Technol China, 96 Jinzhai Rd, Hefei, Anhui, Peoples R China
推荐引用方式 GB/T 7714	He, Li,Zheng, Shuxin,Chen, Wei,et al. OptQuant: Distributed training of neural networks with optimized quantization mechanisms[J]. NEUROCOMPUTING,2019,340:233-244.
APA	He, Li,Zheng, Shuxin,Chen, Wei,Ma, Zhi-Ming,&Liu, Tie-Yan.(2019).OptQuant: Distributed training of neural networks with optimized quantization mechanisms.NEUROCOMPUTING,340,233-244.
MLA	He, Li,et al."OptQuant: Distributed training of neural networks with optimized quantization mechanisms".NEUROCOMPUTING 340(2019):233-244.