面向边缘计算应用的宽度孪生网络

李逸楷; 张通; 陈俊龙

doi:10.16383/j.aas.c200555

面向边缘计算应用的宽度孪生网络

doi: 10.16383/j.aas.c200555

李逸楷¹,
张通^{1, 2},
陈俊龙^{1, 2}

1.
华南理工大学计算机科学与工程学院广州 510006
2.
琶洲实验室广州 510355

基金项目: 国家自然科学基金(61702195, 61751202, U1813203, U1801262, 61751205), 国家科技部重点专项(2019YFA0706200, 2019YFB1703600), 广州市重大科技项项目(202007030006)资助

详细信息

作者简介:
李逸楷：华南理工大学计算机科学与技术专业博士研究生. 2020年获得中国广州华南理工大学计算机科学与技术专业学士学位. 主要研究方向为神经网络, 情感计算等机器学习技术及其在边缘计算领域的应用

张通：华南理工大学计算机科学与工程学院副教授. 2009年获得广州中山大学软件工程专业学士学位, 2011年获得澳门大学应用数学专业硕士学位, 2016年获得澳门大学软件工程专业博士学位. 主要研究方向为情感计算,进化计算,神经网络和其他机器学习技术及其应用. 本文通信作者

陈俊龙：华南理工大学计算机科学与工程学院特聘讲席教授及院长. IEEE Fellow, 美国科学促进会AAAS Fellow, IAPR Fellow, 国际系统及控制论科学院IASCYS院士, 香港工程师学会Fellow, 中国自动化学会Fellow. 欧洲科学院(AE)外籍院士、欧洲科学与艺术学院(EASA)院士, 国际系统与控制论科学院(IASCYS)院士. 1985年毕业于美国密歇根州安娜堡市的密歇根大学安娜堡分校, 2016年获得由母校普渡大学(1988年获得博士学位)颁发的杰出电气和计算机工程师奖. 作为美国工程技术教育认证委员会(ABET)的项目评估员, 成功地设计了澳门大学的工程和计算机科学课程, 并通过香港工程师学会(HKIE)获得了华盛顿/首尔协议的认证. 这被认为是他作为澳门大学科技学院前任院长在工程/计算机科学教育方面的最大贡献. 他在2018年获IEEE系统人机控制论的最高学术奖——IEEE诺伯特•维纳奖(Norbert Wiener Award). 2018和2019年连续两年入选科瑞唯安(Clarivate)全球高被引科学家. 在2012年至2013年, 他是IEEE系统、人和控制论协会主席, 是IEEE期刊《系统、人和控制论: 系统》 (2014-2019)主编. 目前, 他是IEEE控制论汇刊的主编, 以及IEEE模糊系统汇刊的副主编. 他于2015年至2017年任国际自动控制联合会TC9.1经济与商务系统主席. 主要研究方向为控制论、系统和计算智能

计量
- 文章访问数: 1697
- HTML全文浏览量: 429
- PDF下载量: 572
- 被引次数: 0
出版历程
- 收稿日期: 2020-07-15
- 录用日期: 2020-08-27
- 修回日期: 2020-08-18
- 刊出日期: 2020-10-29

Broad Siamese Network for Edge Computing Applications

1.
School of Computer Science and Engineering, South China University of Technology, Guangzhou 510006
2.
Pazhou Laboratory, Guangzhou 510335

Funds: Supported by National Natural Science Foundation of China (61702195, 61751202, U1813203, U1801262, 61751205), National Key Research and Development Program of China (2019YFA0706200, 2019YFB1703600), Science and Technology Major Project of Guangzhou (202007030006)

摘要

摘要: 边缘计算是将计算、存储、通信等任务分配到网络边缘的计算模式. 它强调在用户终端附近执行数据处理过程, 以达到降低延迟, 减少能耗, 保护用户隐私等目的. 然而网络边缘的计算、存储、能源资源有限, 这给边缘计算应用的推广带来了新的挑战. 随着边缘智能的兴起, 人们更希望将边缘计算应用与人工智能技术结合起来, 为我们的生活带来更多的便利. 许多人工智能方法, 如传统的深度学习方法, 需要消耗大量的计算、存储资源, 并且伴随着巨大的时间开销. 这不利于强调低延迟的边缘计算应用的推广. 为了解决这个问题, 我们提出将宽度学习系统(Broad learning system, BLS)等浅层网络方法应用到边缘计算应用领域, 并且设计了一种宽度孪生网络算法. 我们将宽度学习系统与孪生网络结合起来用于解决分类问题. 实验结果表明我们的方法能够在取得与传统深度学习方法相似精度的情况下降低时间和资源开销, 从而更好地提高边缘计算应用的性能.
- 宽度学习系统 /
- 边缘计算 /
- 孪生网络 /
- 浅层网络 /
- 边缘智能
Abstract: Edge computing is a paradigm that allocates computing, storage, communication tasks to the edge of network. It emphasizes that data processing should be placed in the proximity of user terminals in order to reduce latency and energy consumption while protecting user privacy. However, resources for computation, storage and power at the edge of network are limited, which brings new challenges to edge computing applications. With the emergence of edge intelligence, people prefer to combine edge computing applications with artificial intelligence techniques to bring more convenience to our lives. Many artificial intelligence methods such as traditional deep learning methods, need to consume a lot of computation and storage resources with a huge time consumption. It is not conducive to the popularity of edge computing applications, which always require low latency. In order to resolve such a problem, we propose that shallow network methods such as broad learning system can be applied in edge computing applications and design a broad Siamese network algorithm. We combine broad learning system (BLS) with Siamese network for classification tasks. Experimental results show that our method can reduce the cost of time and resources, while achieving a similar result as deep learning methods, consequently improving the performance of edge computing applications.
- Broad learning system (BLS) /
- edge computing /
- Siamese network /
- shallow network /
- edge intelligence

HTML全文

图 1 一种典型的宽度学习系统网络结构

Fig. 1 A typical network structure of broad learning system (BLS)

下载: 全尺寸图片幻灯片

图 2 一种典型的孪生网络结构

Fig. 2 A typical network structure of Siamese network

下载: 全尺寸图片幻灯片

图 3 采用独热编码的相似性度量

Fig. 3 Similarity metrics of one-hot coding

下载: 全尺寸图片幻灯片

图 4 特征映射结果分析

Fig. 4 An analysis of feature mapping results

下载: 全尺寸图片幻灯片

图 5 一种可能的混合特征映射方案

Fig. 5 A possible hybrid feature mapping scheme

下载: 全尺寸图片幻灯片

图 6 宽度孪生网络特征映射过程

Fig. 6 Feature mapping of broad Siamese network

下载: 全尺寸图片幻灯片

图 7 一种基于独热编码的相似度度量方案

Fig. 7 An one-hot based similarity metric

下载: 全尺寸图片幻灯片

图 8 算法准确度随阈值变化的曲线

Fig. 8 Threshold curves of algorithms

下载: 全尺寸图片幻灯片

图 9 受试者工作特征(ROC)曲线

Fig. 9 Receiver operating characteristic (ROC) curves of algorithms

下载: 全尺寸图片幻灯片

图 10 采用不同相似性度量指标宽度孪生网络算法准确度随阈值变化的曲线

Fig. 10 Threshold curves of broad Siamese network with different similarity metrics

下载: 全尺寸图片幻灯片

图 11 采用不同相似性度量指标宽度孪生王洛算法的受试者工作特征(ROC)曲线

Fig. 11 Receiver operating characteristic (ROC) curves of broad Siamese network with different similarity metrics

下载: 全尺寸图片幻灯片

表 1 实验数据集信息表

Table 1 Table of data set for experiments

数据集	样本规模	类别数量	各类别样本数量	特征维度
CK+	5876	7	(1022, 233, 868, 546, 1331, 547, 1329)	14400
MNIST	70000	10	每个类别近似 7000 样本	784
JAFFE	213	7	(30, 29, 32, 31, 30, 31, 30)	14 400
USPS	20000	10	每个类别 2 000 样本	784

下载: 导出CSV

表 2 宽度孪生网络参数设置信息

Table 2 Table of parameters for broad Siamese network

数据集	n	p	e
CK+	8	10	9000
MNIST	10	10	500
JAFFE	8	10	9000
USPS	10	10	1500

下载: 导出CSV

表 3 对比算法中全连接神经网络结点个数设置信息

Table 3 Table of number about nodes in the fully connected network for comparison

数据集	第一层结点数	第二层结点数	第三层结点数
CK+	512	128	512
MNIST	16	16	16
JAFFE	1024	128	1024
USPS	128	128	128

下载: 导出CSV

表 4 准确率实验结果

Table 4 Table of experiment results about accuracy

数据集	宽度孪生网络	基于全连接神经网络的孪生网络
CK+	0.9788738	0.9287094
MNIST	0.9798112	0.9777414
JAFFE	0.9217687	0.9206349
USPS	0.9536075	0.9505025

下载: 导出CSV

表 5 训练时间实验结果

Table 5 Table of experiment results about training time

数据集	宽度孪生网络	基于全连接神经网络的孪生网络
CK+	94.140997	567.9896214
MNIST	5.6314652	60.0518959
JAFFE	58.4795067	1105.1385579
USPS	3.0834677	29.8392925

下载: 导出CSV

表 6 内存开销实验结果

Table 6 Table of experiment results about memory overhead

数据集	宽度孪生网络	基于全连接神经网络的孪生网络
CK+	3.3491211	6.0307884
MNIST	2.4598732	2.0554810
JAFFE	0.2893066	5.6162262
USPS	1.2569504	0.8804893

下载: 导出CSV

参考文献(16)

[1]	Chen C L P , Liu Z. Broad learning system: an effective and efficient incremental learning system without the need for deep architecture. IEEE Transactions on Neural Networks and Learning Systems, 2018, 29(1): 10−24 doi: 10.1109/TNNLS.2017.2716952
[2]	Peng X, Ota K, Dong M. A broad learning-driven network traffic analysis system based on fog computing paradigm. China Communications, 2020, 17(2): 1−13 doi: 10.23919/JCC.2020.02.001
[3]	Zhang T, Liu Z, Wang X, Xing X, Chen C L P, Chen E. Facial expression recognition via broad learning system. In: Proceedings of the 2018 IEEE International Conference on Systems, Man, and Cybernetics (SMC), Miyazaki, Japan: IEEE, 2018. 1898−1902
[4]	Gao S, Guo G, Huang H, Cheng X, Chen C L P. An end-to-end broad learning system for event-based object classification. IEEE Access, 2020, 8: 45974−45984 doi: 10.1109/ACCESS.2020.2978109
[5]	Liu X, Qiu T, Chen C, Ning H, Chen N. An incremental broad learning approach for semi-supervised classification. In: Prceedings of the 2019 IEEE International conference on Dependable, Autonomic and Secure Computing, International conference on Pervasive Intelligence and Computing, International conference on Cloud and Big Data Computing, International conference on Cyber Science and Technology Congress (DASC/PiCom/CBDCom/CyberSciTech), Fukuoka, Japan: IEEE. 2019. 250−254
[6]	Wang X, Zhang T, Xu X, Chen L, Xing X, Chen C L P. EEG emotion recognition using dynamical graph convolutional neural networks and broad learning system. In: Proceedings of the 2018 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), Madrid, Spain: 2018.1240−1244
[7]	Chu F, Liang T, Chen C L P, Wang X, Ma X. Weighted broad learning system and its application in nonlinear industrial process modeling. IEEE Transactions on Neural Networks and Learning Systems. 2020, 31(8): 3017−3031
[8]	Bromley J, Guyon I, Lecun Y, et al. Signature verification using a Siamese time delay neural network. In: Proceedings of the Advances in Neural Information Processing Systems 6, 7th NIPS Conference, Denver, Colorado, USA: Morgan Kaufmann Publishers Inc., 1993.
[9]	Nair V, Hinton G E. Rectified linear units improve restricted boltzmann machines. In: Proceedings of the 27th International Conference on International Conference on Machine Learning (ICML´10). Omnipress, Madison, WI, USA: 2019.807−814
[10]	Treible W, Saponaro P, Kambhamettu C. Wildcat: in-the-wild color-and-thermal patch comparison with deep residual pseudo-Siamese networks. In: Proceedings of the 2019 IEEE International Conference on Image Processing (ICIP), Taipei, China: IEEE, 2019. 1307−1311
[11]	Baraldi L, Grana C, Cucchiara R. A deep Siamese network for scene detection in broadcast videos. In: Proceedings of the 23rd ACM International Conference on Multimedia (MM'15). Association for Computing Machinery, New York, NY, USA: 1199−1202
[12]	Melekhov I, Kannala J, Rahtu E. Siamese network features for image matching. In: Proceedings of the 23rd International Conference on Pattern Recognition (ICPR), Cancun, Mexico: 2016. 378−383
[13]	Bertinetto L , Valmadre J , Henriques J F , et al. Fully-convolutional Siamese networks for object tracking. In: Proceedings of European Conference on Computer Vision (ECCV) Workshops. 2016. 850—865
[14]	Zeghidour N, Synnaeve G, Usunier N, et al. Joint learning of speaker and phonetic similarities with siamese networks. In: Proceedings of Interspeech 2016, 2016. 1295−1299
[15]	Neculoiu P, Versteegh M, Rotaru M. Learning text similarity with Siamese recurrent networks. In: Proceedings of the 1st Workshop on Representation Learning for NLP. 2016. 148−157
[16]	Rahul M V, Ambareesh R, Shobha G. Siamese network for underwater multiple object tracking. In: Proceedings of the 9th International Conference on Machine Learning and Computing (ICMLC 2017). New York, NY, USA: ACM, 2017. 511−516