| Presentation | 2017-11-07 DCNN Training with Short Bit Length Format Considering Loss of Trailing Digits Shin-ichi O'uchi, Hiroshi Fuketa, Ryousei Takano, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | Loss of trailing digits in training deep convolutional neural network (DCNN) is considered to implement training with a format shorter than 16-bit floating format. As a result, we propose an 8-bit training method. The method improves inference accuracy compared to 16-bit floating format in training AlexNet. Simultaneously, the hardware size is reduced by half compared to the 16-bit hardware. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | Deep convolutional neural network (DCNN) / AlexNet / training / floating point arithmetic / accuracy |
| Paper # | CPSY2017-41 |
| Date of Issue | 2017-10-31 (CPSY) |
| Conference Information | |
| Committee | VLD / DC / CPSY / RECONF / CPM / ICD / IE / IPSJ-SLDM / IPSJ-EMB / IPSJ-ARC |
|---|---|
| Conference Date | 2017/11/6(3days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | Kumamoto-Kenminkouryukan Parea |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | Design Gaia 2017 -New Field of VLSI Design- |
| Chair | Hiroyuki Ochi(Ritsumeikan Univ.) / Michiko Inoue(NAIST) / Koji Nakano(Hiroshima Univ.) / Masato Motomura(Hokkaido Univ.) / Fumihiko Hirose(Yamagata Univ.) / Hideto Hidaka(Renesas) / Takayuki Hamamoto(Tokyo Univ. of Science) / Kiyoharu Hamaguchi(Shimane Univ.) / 渡辺 晴美(東海大) / Masahiro Goshima(NII) |
| Vice Chair | Noriyuki Minegishi(Mitsubishi Electric) / Satoshi Fukumoto(Tokyo Metropolitan Univ.) / Hidetsugu Irie(Univ. of Tokyo) / Takashi Miyoshi(Fujitsu) / Yuichiro Shibata(Nagasaki Univ.) / Kentaro Sano(Tohoku Univ.) / Mayumi Takeyama(Kitami Inst. of Tech.) / Makoto Nagata(Kobe Univ.) / Kazuya Kodama(NII) / Hideaki Kimata(NTT) |
| Secretary | Noriyuki Minegishi(Hiroshima City Univ.) / Satoshi Fukumoto(NTT) / Hidetsugu Irie(Kyoto Sangyo Univ.) / Takashi Miyoshi(Tokyo Inst. of Tech.) / Yuichiro Shibata(Utsunomiya Univ.) / Kentaro Sano(Hokkaido Univ.) / Mayumi Takeyama(Hiroshima City Univ.) / Makoto Nagata(e-trees.Japan) / Kazuya Kodama(Nihon Univ.) / Hideaki Kimata(Toyohashi Univ. of Tech.) / (Univ. of Tokyo) / (Panasonic) / (Nagoya Univ.) |
| Assistant | / Masayuki Arai(Nihon Univ.) / Yasuaki Ito(Hiroshima Univ.) / Tomoaki Tsumura(Nagoya Inst. of Tech.) / Yuuki Kobayashi(NEC) / Hiroki Nakahara(Tokyo Inst. of Tech.) / Yuichi Akage(NTT) / Masanori Natsui(Tohoku Univ.) / Masatoshi Tsuge(Socionext) / Hiroyuki Ito(Tokyo Inst. of Tech.) / Pham Konkuha(Univ. of Electro-Comm.) / Yasutaka Matsuo(NHK) / Kazuya Hayase(NTT) |
| Paper Information | |
| Registration To | Technical Committee on VLSI Design Technologies / Technical Committee on Dependable Computing / Technical Committee on Computer Systems / Technical Committee on Reconfigurable Systems / Technical Committee on Component Parts and Materials / Technical Committee on Integrated Circuits and Devices / Technical Committee on Image Engineering / 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) | DCNN Training with Short Bit Length Format Considering Loss of Trailing Digits |
| Sub Title (in English) | |
| Keyword(1) | Deep convolutional neural network (DCNN) |
| Keyword(2) | AlexNet |
| Keyword(3) | training |
| Keyword(4) | floating point arithmetic |
| Keyword(5) | accuracy |
| 1st Author's Name | Shin-ichi O'uchi |
| 1st Author's Affiliation | National Institute of AIST(AIST) |
| 2nd Author's Name | Hiroshi Fuketa |
| 2nd Author's Affiliation | National Institute of AIST(AIST) |
| 3rd Author's Name | Ryousei Takano |
| 3rd Author's Affiliation | National Institute of AIST(AIST) |
| Date | 2017-11-07 |
| Paper # | CPSY2017-41 |
| Volume (vol) | vol.117 |
| Number (no) | CPSY-278 |
| Page | pp.pp.7-12(CPSY), |
| #Pages | 6 |
| Date of Issue | 2017-10-31 (CPSY) |