Presentation | 2024-03-13 A Plug-and-Play Module for Enhancing Fault-Tolerant Distributed Inference Based on Gaussian Dropout Hou Zhangcheng, Ohtsuki Tomoaki, |
---|---|
PDF Download Page | PDF download Page Link |
Abstract(in Japanese) | (See Japanese page) |
Abstract(in English) | Distributed inference (DI) in the Internet of Things (IoT) is becoming increasingly important as the demand for AI applications grows. When unreliable links are used for IoT transmission, missing data can adversely affect inference accuracy. Therefore, we designed a plug-and-play module to enhance the robustness of DI systems. The dropout we utilize is no longer a fixed value but is sampled from a Gaussian distribution, making it better adapted to lossy networks. Additionally, the convolutional layer of the module can learn and preserve fault-tolerant inference. This method uses freeze training and does not require retraining of the original deep neural network, thus significantly reducing the number of parameters to be trained and avoiding catastrophic forgetting. We incorporate various designs for the Gaussian dropout layer to further enhance its effectiveness at low packet loss rates. The module only needs to be attached to the cut-points of the original deep neural network, making it very easy and fast to deploy. Experimental results demonstrate that our plug-and-play module can be adapted to different deep neural networks and various lossy network scenarios, achieving an average accuracy improvement in all tests. Results show that our plug-and-play module significantly enhances the fault-tolerant inference of the system. |
Keyword(in Japanese) | (See Japanese page) |
Keyword(in English) | Internet of Thingserror-tolerantdistributed inferencedeep learning |
Paper # | RCS2023-267 |
Date of Issue | 2024-03-06 (RCS) |
Conference Information | |
Committee | RCS / SR / SRW |
---|---|
Conference Date | 2024/3/13(3days) |
Place (in Japanese) | (See Japanese page) |
Place (in English) | The University of Tokyo (Hongo Campus), and online |
Topics (in Japanese) | (See Japanese page) |
Topics (in English) | Mobile Communication Workshop |
Chair | Kenichi Higuchi(Tokyo Univ. of Science) / Osamu Takyu(Shinshu Univ.) / Keiichi Mizutani(Kyoto Univ.) |
Vice Chair | Fumihide Kojima(NICT) / Osamu Muta(Kyushu Univ.) / Naoto Ishii(NEC) / Kentaro Ishidu(NICT) / Kazuto Yano(ATR) / Shusuke Narieda(Mie Univ.) / Kentaro Saito(Tokyo Denki Univ.) / Hirokazu Sawada(NICT) |
Secretary | Fumihide Kojima(Univ. of Electro-Comm) / Osamu Muta(Sharp) / Naoto Ishii(Mitsubishi Electric) / Kentaro Ishidu(Tokai Univ.) / Kazuto Yano(NTT) / Shusuke Narieda(Tokyo Inst. of Tech) / Kentaro Saito(Kochi Univ, of Tech.) / Hirokazu Sawada(Niigata Univ.) |
Assistant | Masashi Iwabuchi(NTT) / Issei Kanno(KDDI Research) / Yuyuan Chang(Tokyo Inst. of Tech) / Kazuki Maruta(Tokyo Univ. of Science) / Kiichi Tateishi(NTT Docomo) / Taichi Ohtsuji(NEC) / WANG Xiaoyan(Ibaraki Univ.) / Akemi Tanaka(MathWorks) / Katsuya Suto(Univ. of Electro-Comm) / Maki Arai(Tokyo Univ. of Science) / Yusuke Kouda(Kyoto Univ.) |
Paper Information | |
Registration To | Technical Committee on Radio Communication Systems / Technical Committee on Smart Radio / Technical Committee on Short Range Wireless Communications |
---|---|
Language | ENG |
Title (in Japanese) | (See Japanese page) |
Sub Title (in Japanese) | (See Japanese page) |
Title (in English) | A Plug-and-Play Module for Enhancing Fault-Tolerant Distributed Inference Based on Gaussian Dropout |
Sub Title (in English) | |
Keyword(1) | Internet of Thingserror-tolerantdistributed inferencedeep learning |
1st Author's Name | Hou Zhangcheng |
1st Author's Affiliation | Keio University(KU) |
2nd Author's Name | Ohtsuki Tomoaki |
2nd Author's Affiliation | Keio University(KU) |
Date | 2024-03-13 |
Paper # | RCS2023-267 |
Volume (vol) | vol.123 |
Number (no) | RCS-434 |
Page | pp.pp.77-82(RCS), |
#Pages | 6 |
Date of Issue | 2024-03-06 (RCS) |