南京大学：《并行处理技术——分布式与并行计算 Distributed and Parallel computing（并行计算——结构、算法、编程）》课程教学资源（课件讲稿）专题三 边缘智能（边缘计算时代的人工智能）

边缘智能 边缘计算时代的人工智能 报告人：谢磊 南京大学 2020-3-15

www.educity.cn —  
# ƊŐûŘ ƊŐūŃ÷8ī1²ûŘ à‚1Ŷ ķ n0Ÿ 

提纲 边缘智能简介 基于边缘智能的视频大数据分析 案例：基于视频分析的抬头检测系统

www.educity.cn —  
# éň n ƊŐûŘł4 n +ƊŐûŘīŧƮðåZĄ n ċH+ŧƮZĄīáČėņŏ

提纲 ■边缘智能简介 ■基于边缘智能的视频大数据分析 ■案例：基于视频分析的抬头检测系统

www.educity.cn —  
# éň n ƊŐûŘł4 n +ƊŐûŘīŧƮðåZĄ n ċH+ŧƮZĄīáČėņŏ

“大数据”的产生：云→边缘 All Data Created by Category 2011-2020(Zetabytes) 160.0 140.0 120.0 1000 800 600 Edge 40.0 Data 20.0 20112012 2013 201420152016 2017201820192020 Consumer Images,Voice Video 01 02 03 04 0.6 10 15 23 35 54 Entertainment Social Media 03 05 07 11 17 27 4.3 67 104 162 Data Processing 0.4 0.6 10 17 28 46 76 12.5 205 337 Medical 0.0 01 01 02 04 0.7 12 2.0 35 61 internet of Things 01 02 03 05 08 14 23 40 6水 115 Surveilance 09 14 23 17 59 95 15.1 242 林游 620 rce:Wbon 2015 based on wbon 2013 Prerectons 预计在2020年，边缘数据将达到80%

ðåī/Ħ, à ƊŐ ƭūŠ2020¼ƊŐðå«Ƌ_80%

为什么需要边缘智能(Edge Intelligence)? ·如在云端处理大数据，会产生带宽、时延问题 ·基于大数据的智能应用（如自动驾驶、AR等）需要实 时处理 ·利用云计算的算力集中式处理，会带来带宽、时延 的巨大开销 ·需要把A能力推送边缘，构成边缘智能(EI) ·近端感知数据中的变化，以响应快速变化的环境， 提高操作效率 Edge Intelligence ·边缘节点本地处理，避免原始数据传输带来的网络 Edge Computing Enablers 负担 08g Edge Storage

%2'ƪťƊŐûŘ(Edge Intelligence) • ™Š,ŀĥðå?/Ħ¹§÷ÂƢƯ • +ðåīûŘ¿ħ(™ŚfƲƱARŁ)ƪť¤ ÷ĥ • ]ħ,ūŃīŃcƩ#Äĥ?¹Ă¹§÷ ī³ÃƟ • ƪťßAIŘcèƔƊŐă×ƊŐûŘ(EI) • ƏŀÖijðå#īyj9…¿ÏƗyjīģ‘ éƳëGíĠ • ƊŐŜěÿ‹ĥƙPtšðå@ƈ¹ĂīőŎ ŷâ

边缘智能(Edge Intelligence) Cloud data center 。云-边-端协同处理框架 Deep learning ●将人工智能算法（当前主要是 Internet Deep 深度学习)推向边缘 learning 九 Edge Edge ●加速AI的发展 server server ·驱动力有四个方面：算法、硬件、 Edge devices 数据和应用场景 End devices ● 边缘端丰富的应用场景和多样化 Deep learning 的数据促进了AI进一步发展 [1]JIASI CHEN AND XUKAN RAN,Deep Learning With Edge Computing:A Review,Proceedings of the IEEE,May 2019

ƊŐûŘ(Edge Intelligence) [1] JIASI CHEN AND XUKAN RAN, Deep Learning With Edge Computing: A Review, Proceedings of the IEEE, May 2019. l ,-Ɗ-ŀm}ĥĊĆ l «1²ûŘŃĕÉa&ťø ęÁŸ)è~ƊŐ l dƗAIīw¯ l ưfcþ‡"ôƬŃĕĵ; ð儿ħŒú l ƊŐŀ$¨ī¿ħŒú„Ĉj īðåLƐ*AIƐĐw¯

深度学习的训练和推理 Cloud ● 经典模式：云端训练+边缘推理 Cloud Training (Takes hours,Days) Forward ·选配模式：边缘训练，基于本 地的数据特性来进行定制化的 Cloud Data 训练 Edge ● 通过Pruning、Dropout等方 Edge Data Generation Edge Inference (Takes ms.seconds) 法降低训练参数量 ●通过迁移学习快速训练 Edge Local Training ·通过强化学习迭代训练

ęÁŸ)īŭʼn„èĥ l ŊUčÄ,ŀŭʼnƊŐèĥ l ƕƜčÄƊŐŭʼn+ÿ ‹īðåĞÓĂƐŢ¢`jī ŭʼn l ƖƍPruningDropoutŁô ĕƦDŭʼnuðƞ l ƖƍƌļŸ)ÏƗŭʼn l ƖƍÈjŸ)ƒ8ŭʼn

边缘训练模型 Edge server DNN to be Gradient update trained Gradient can be compressed or sen infrequently Gradient update End device End device (b)Decentralized training. (a)Centralized training. ·集中式训练：Each worker computes local data set,which are then collected by a central parameter server,and the updates sent back to the workers.代表性模型：联邦学习 ·分布式训练：Each device computes its own gradient updates based on its training data and then communicates its updates to some of the other devices. ·两种模型比较：DNN模型一致性、带宽限制等

ƊŐŭʼnčŽ • Ʃ#Äŭʼn: Each worker computes local data set, which are then collected by a central parameter server, and the updates sent back to the workers. 8ţÓčŽŖƚŸ) • Z·Äŭʼn: Each device computes its own gradient updates based on its training data and then communicates its updates to some of the other devices. • !ĻčŽĒƆDNNčŽśÓ¹§Ƨ`Ł

边缘推理模型 Edge Edge Server Server (a)Edge-based mode Device A Device B (b)Device-based mode (a)Edge-based mode. (b)Device-based mode. (c)Edge-device mode Edge Edge Server Server (d)Edge-cloud mode Cloud Data Center Device Device D (c)Edge-device mode. (d)Edge-cloud mode

ƊŐèĥčŽ
aEdge-based mode
bDevice-based mode
cEdge-device mode
dEdge-cloud mode

边缘智能的6个等级 Level 6 All on-device 根据数据卸载的数量和路 Level 5 All in-adge 径长度，我们将边缘智能 Reduced amount or shorten path of data offloading Levol 4 分成6个等级 Cloud-edge co-training Level3 On-device inference 层级越高，数据卸载的数 量和路径长度越少，从而 Lovel 2 Training In-edge co-inference on the cloud 时延、带宽、安全有优势， Level 1 但会增加计算延迟和能耗 Cloud-edge co-inference 代价 Cloud Intelligence Training and inference on the cloud [2]Zhi Zhou,Xu Chen,En Li,Liekang Zeng,Ke Luo,and Junshan Zhang,Edge Intelligence:Paving the Last Mile of Artificial Intelligence with Edge Computing,Proceedings of the IEEE,May 2019

ƊŐûŘī"ŁŇ [2] Zhi Zhou, Xu Chen, En Li, Liekang Zeng, Ke Luo, and Junshan Zhang, Edge Intelligence: Paving the Last Mile of Artificial Intelligence with Edge Computing, Proceedings of the IEEE, May 2019. • ĉåðåsƅīðƞ„ſ ÍơÁØ:«ƊŐûŘ Z×"ŁŇ • ®ŇżƳðåsƅīð ƞ„ſÍơÁż­6Ŕ ÷¹§ Rþ>h B?’dūŃÂƑ„Řŕ 8<