Presentation | 2022-01-24 Addition of DPU Training Function by Tail Layer Training Yuki Takashima, Akira Jinguji, Hiroki Nakahara, |
---|---|
PDF Download Page | PDF download Page Link |
Abstract(in Japanese) | (See Japanese page) |
Abstract(in English) | The demand for deep learning has been increasing, and many hardware implementations have been proposed. The Deep learning Processor Unit (DPU) was provided by Xilinx. Althogh it can perform inference at high speed, it cannot perform training. We propose a tail layer training that makes the tail layer of a Convolutional Neural Network (CNN) independent. All layers except the tail layer are computed using a DPU, and the tail layer is computed by a CPU. Since the number of neurons and classes in the output must be the same for image classification, it is effective for retraining to add the number of classes. The tail layer training, found that the relationship between the existing classes and the classes to be added is important. Therefore, it is not suitable for training on large number of classes. However, with a dataset such as cifar10, it is able to reduce the loss of accuracy by about 3 points between training the entire model with all 10 classes and training only the tail layer with 2 add classes after training the entire model with 8 classes. |
Keyword(in Japanese) | (See Japanese page) |
Keyword(in English) | CNN / Image Classification / DPU / Tail Layer Training |
Paper # | VLD2021-59,CPSY2021-28,RECONF2021-67 |
Date of Issue | 2022-01-17 (VLD, CPSY, RECONF) |
Conference Information | |
Committee | RECONF / VLD / CPSY / IPSJ-ARC / IPSJ-SLDM |
---|---|
Conference Date | 2022/1/24(2days) |
Place (in Japanese) | (See Japanese page) |
Place (in English) | Online |
Topics (in Japanese) | (See Japanese page) |
Topics (in English) | FPGA Applications, etc. |
Chair | Kentaro Sano(RIKEN) / Kazutoshi Kobayashi(Kyoto Inst. of Tech.) / Michihiro Koibuchi(NII) / Hiroshi Inoue(Kyushu Univ.) / Yuichi Nakamura(NEC) |
Vice Chair | Yoshiki Yamaguchi(Tsukuba Univ.) / Tomonori Izumi(Ritsumeikan Univ.) / Minako Ikeda(NTT) / Kota Nakajima(Fujitsu Lab.) / Tomoaki Tsumura(Nagoya Inst. of Tech.) |
Secretary | Yoshiki Yamaguchi(NEC) / Tomonori Izumi(Tokyo Inst. of Tech.) / Minako Ikeda(Osaka Univ.) / Kota Nakajima(NEC) / Tomoaki Tsumura(JAIST) / (Hitachi) / (Univ. of Tokyo) |
Assistant | Yukitaka Takemura(INTEL) / Yasunori Osana(Ryukyu Univ.) / / Ryohei Kobayashi(Tsukuba Univ.) / Takaaki Miyajima(Meiji Univ.) |
Paper Information | |
Registration To | Technical Committee on Reconfigurable Systems / Technical Committee on VLSI Design Technologies / Technical Committee on Computer Systems / Special Interest Group on System Architecture / Special Interest Group on System and LSI Design Methodology |
---|---|
Language | JPN |
Title (in Japanese) | (See Japanese page) |
Sub Title (in Japanese) | (See Japanese page) |
Title (in English) | Addition of DPU Training Function by Tail Layer Training |
Sub Title (in English) | |
Keyword(1) | CNN |
Keyword(2) | Image Classification |
Keyword(3) | DPU |
Keyword(4) | Tail Layer Training |
1st Author's Name | Yuki Takashima |
1st Author's Affiliation | Tokyo Institute of Technology(Tokyo Tech) |
2nd Author's Name | Akira Jinguji |
2nd Author's Affiliation | Tokyo Institute of Technology(Tokyo Tech) |
3rd Author's Name | Hiroki Nakahara |
3rd Author's Affiliation | Tokyo Institute of Technology(Tokyo Tech) |
Date | 2022-01-24 |
Paper # | VLD2021-59,CPSY2021-28,RECONF2021-67 |
Volume (vol) | vol.121 |
Number (no) | VLD-342,CPSY-343,RECONF-344 |
Page | pp.pp.55-60(VLD), pp.55-60(CPSY), pp.55-60(RECONF), |
#Pages | 6 |
Date of Issue | 2022-01-17 (VLD, CPSY, RECONF) |