| Presentation | 2018-03-08 A Study of Kernel Clustering for Reducing Memory Footprint of CNN Yuki Matsui, Shinobu Miwa, Satoshi Shindo, Tomoaki Tsumura, Hayato Yamaki, Hiroki Honda, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | Convolutional Neural Network (CNN) is widely used in the field of image recognition due to the high recognition accuracy. CNN is a sort of deep and large-scale neural networks so that it has numbers of parameters to be used for the computation. There have been many studies of compressing the data of CNN such as reducing the numbers of parameters and bits of parameters. Meanwhile, a well-trained CNN has very regular structure (i.e., 2D kernels) available for data compression, but no study of exploiting this structure for data compression in CNN has been reported so far. We have proposed a technique that clusters 2D kernels trained and replaces them with representative 2D kernels for reducing the number of parameters in CNN. In this paper, we report the experimental results of clustering the overall 2D kernels within VGG-16 with various numbers of clusters. Our experimental results show that the proposed technique can reduce the number of kernels by 85.6% in exchange for a 9% reduction in the recognition accuracy. The proposed technique is orthogonal to the other approaches of compressing the data in CNN, such as pruning and quantization; hence, they can be used together to obtain further gains. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | CNN / data compression / clustering |
| Paper # | CPSY2017-140,DC2017-96 |
| Date of Issue | 2018-02-28 (CPSY, DC) |
| Conference Information | |
| Committee | CPSY / DC / IPSJ-SLDM / IPSJ-EMB / IPSJ-ARC |
|---|---|
| Conference Date | 2018/3/7(2days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | Okinoshima Bunka-Kaikan Bldg. |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | ETNET2018 |
| Chair | Koji Nakano(Hiroshima Univ.) / Michiko Inoue(NAIST) / Kiyoharu Hamaguchi(Shimane Univ.) / / Masahiro Goshima(NII) |
| Vice Chair | Hidetsugu Irie(Univ. of Tokyo) / Takashi Miyoshi(Fujitsu) / Satoshi Fukumoto(Tokyo Metropolitan Univ.) |
| Secretary | Hidetsugu Irie(Utsunomiya Univ.) / Takashi Miyoshi(Hokkaido Univ.) / Satoshi Fukumoto(Kyoto Sangyo Univ.) / (Tokyo Inst. of Tech.) / (Panasonic) / (Kochi Univ. of Tech.) |
| Assistant | Yasuaki Ito(Hiroshima Univ.) / Tomoaki Tsumura(Nagoya Inst. of Tech.) / Masayuki Arai(Nihon Univ.) |
| Paper Information | |
| Registration To | Technical Committee on Computer Systems / Technical Committee on Dependable Computing / Special Interest Group on System and LSI Design Methodology / Special Interest Group on Embedded Systems / Special Interest Group on System Architecture |
|---|---|
| Language | JPN |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | A Study of Kernel Clustering for Reducing Memory Footprint of CNN |
| Sub Title (in English) | |
| Keyword(1) | CNN |
| Keyword(2) | data compression |
| Keyword(3) | clustering |
| 1st Author's Name | Yuki Matsui |
| 1st Author's Affiliation | The University of Electro-Communications(UEC) |
| 2nd Author's Name | Shinobu Miwa |
| 2nd Author's Affiliation | The University of Electro-Communications(UEC) |
| 3rd Author's Name | Satoshi Shindo |
| 3rd Author's Affiliation | Nagoya Institute of Technology(NITech) |
| 4th Author's Name | Tomoaki Tsumura |
| 4th Author's Affiliation | Nagoya Institute of Technology(NITech) |
| 5th Author's Name | Hayato Yamaki |
| 5th Author's Affiliation | The University of Electro-Communications(UEC) |
| 6th Author's Name | Hiroki Honda |
| 6th Author's Affiliation | The University of Electro-Communications(UEC) |
| Date | 2018-03-08 |
| Paper # | CPSY2017-140,DC2017-96 |
| Volume (vol) | vol.117 |
| Number (no) | CPSY-479,DC-480 |
| Page | pp.pp.185-190(CPSY), pp.185-190(DC), |
| #Pages | 6 |
| Date of Issue | 2018-02-28 (CPSY, DC) |