BottleNet: A Deep Learning Architecture for Intelligent Mobile Cloud Computing Services

被引：115
作者：
Eshratifar, Amir Erfan [1 ]
Esmaili, Amirhossein [1 ]
Pedram, Massoud [1 ]
机构：
[1] Univ Southern Calif, Dept Elect & Comp Engn, Los Angeles, CA 90007 USA
来源：
2019 IEEE/ACM INTERNATIONAL SYMPOSIUM ON LOW POWER ELECTRONICS AND DESIGN (ISLPED) | 2019年
关键词：
deep learning; collaborative intelligence; mobile computing; cloud computing; feature compression; NEURAL-NETWORKS; COLLABORATIVE INTELLIGENCE;
D O I：
10.1109/islped.2019.8824955
中图分类号：
TP3 [计算技术、计算机技术];
学科分类号：
0812 ;
摘要：
Recent studies have shown the latency and energy consumption of deep neural networks can be significantly improved by splitting the network between the mobile device and cloud. This paper introduces a new deep learning architecture, called BottleNet, for reducing the feature size needed to be sent to the cloud. Furthermore, we propose a training method for compensating for the potential accuracy loss due to the lossy compression of features before transmitting them to the cloud. BottleNet achieves on average 5.1x improvement in end-to-end latency and 6.9x improvement in mobile energy consumption compared to the cloud-only approach with no accuracy loss.
引用
收藏
页数：6
相关论文
共 22 条
[1]  
[Anonymous], 2017, US SPEEDTEST MARKET
[2]  
[Anonymous], 2018, NVIDIA TensorRT
[3]  
Chen Zhenhua, 2018, ARXIV180808692
[4]  
Chetlur S., 2014, CUDNN EFFICIENT PRIM
[5]  
Choi H., 2018, 2018 6th International Conference on Brain-Computer Interface (BCI), P1
[6]  
Choi H, 2018, IEEE IMAGE PROC, P3743, DOI 10.1109/ICIP.2018.8451100
[7]  
Deng J, 2009, PROC CVPR IEEE, P248, DOI 10.1109/CVPRW.2009.5206848
[8]  
Eshratifar A. E., 2018, ARXIV PREPRINT ARXIV
[9]  
Eshratifar AE, 2019, INT SYM QUAL ELECT, P14, DOI [10.1109/isqed.2019.8697647, 10.1109/ISQED.2019.8697647]
[10]   Energy and Performance Efficient Computation Offloading for Deep Neural Networks in a Mobile Cloud Computing Environment [J].
Eshratifar, Amir Erfan ;
Pedram, Massoud .
PROCEEDINGS OF THE 2018 GREAT LAKES SYMPOSIUM ON VLSI (GLSVLSI'18), 2018, :111-116
123