Deep Convolutional Neural Network on iOS mobile devices (Invited Paper)

Chun Fu Chen, Gwo Giun Lee, Vincent Sritapan, Ching Yung Lin

研究成果: Conference contribution

12 引文 斯高帕斯(Scopus)

摘要

Deep Convolutional Neural Network (CNN) draws significant attention in the computer vision community by facilitating machines with more intelligence in understanding visual signals; however, its computation complexity has also increased significantly. To achieve ubiquitous machine intelligence, deep CNN is required to be ported onto local devices rather than cloud-based solution due to low latency consideration. Hence, in this paper, we propose a method to explore the design space for porting deep CNN onto iOS mobile devices, with attempts in maximizing data reusability, which alleviates the high bandwidth burden in the convolution layers of CNN. Furthermore, effective data reuse also makes possible the parallelization of all computing threads without data loading latency. On the other hand, deep CNN is usually over-parametrized with many unused convolution kernels. Based on Algorithm/Architecture Co-Exploration, we introduced a method in pruning redundant kernels in deep CNN with ignorable performance degradation on validation dataset (0.06% loss). This reduces 29% of operations and 34% of storage size on a 16-layer CNN. We used iPhone 6s and iPad Pro for case studies, and ported 8-layer and 16-layer CNNs onto targeted devices. The data reusability strategy improves computation speed up to 1.3×; and redundant kernel removal increases computation speed to 1.43×. As a result, we achieved high computation efficiency and have thus enhanced the capability of machine intelligence on local mobile devices.

原文English
主出版物標題Proceedings - IEEE International Workshop on Signal Processing Systems, SiPS 2016
發行者Institute of Electrical and Electronics Engineers Inc.
頁面130-135
頁數6
ISBN(電子)9781509033614
DOIs
出版狀態Published - 2016 12月 9
事件2016 IEEE International Workshop on Signal Processing Systems, SiPS 2016 - Dallas, United States
持續時間: 2016 10月 262016 10月 28

出版系列

名字IEEE Workshop on Signal Processing Systems, SiPS: Design and Implementation
ISSN(列印)1520-6130

Other

Other2016 IEEE International Workshop on Signal Processing Systems, SiPS 2016
國家/地區United States
城市Dallas
期間16-10-2616-10-28

All Science Journal Classification (ASJC) codes

  • 電氣與電子工程
  • 訊號處理
  • 應用數學
  • 硬體和架構

指紋

深入研究「Deep Convolutional Neural Network on iOS mobile devices (Invited Paper)」主題。共同形成了獨特的指紋。

引用此