Presentation 2018-08-01
Optimization of Numerical Expression in CNN using Genetic Algorithm
Wakana Nogami, Tsutomu Ikegami, Shin-ichi O'uchi, Ryosei Takano, Yuma Kishi, Tomohiro Kudoh,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) The accuracy of image recognition by the convolutional neural network (CNN) has been improving year by year, and the models are becoming more complicated and larger. One way to reduce the model size is to use a low bit-width numerical expression. There are many types of research to reduce the bit-width by using floating-point, fixed-point, ternary, and binary arithmetics, and so on. They are aiming at (1) simplifying the computation (2) by introducing a less-bit arithmetic (3) to keep the image recognition accuracy. Considering ease of computation, it is appropriate to use floating-point and fixed-point. Therefore, we introduced a variable bin size quantization and found the appropriate numerical expression from the viewpoint of low bit-width and high accuracy by optimizing the bin size using a genetic algorithm. In this study, we targeted on quantization of trained parameters at inference. As a result, we succeeded in finding a numerical expression that can give higher Top-1 Accuracy than when using fixed-point or floating-point type for our CNN models. The numerical expression is relatively similar to fixed-point type. We also found that by using this numerical expression, it is possible to reduce the bit-width down to 3-bit without decreasing the accuracy.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) CNN / Adaptive Quantization / Genetic Algorithm / Numerical Expression / Optimization
Paper # CPSY2018-26
Date of Issue 2018-07-23 (CPSY)

Conference Information
Committee CPSY / DC / IPSJ-ARC
Conference Date 2018/7/30(3days)
Place (in Japanese) (See Japanese page)
Place (in English) Kumamoto City International Center
Topics (in Japanese) (See Japanese page)
Topics (in English) Parallel, Distributed and Cooperative Processing Systems and Dependable Computing
Chair Koji Nakano(Hiroshima Univ.) / Satoshi Fukumoto(Tokyo Metropolitan Univ.) / Masahiro Goshima(NII)
Vice Chair Hidetsugu Irie(Univ. of Tokyo) / Takashi Miyoshi(Fujitsu) / Hiroshi Takahashi(Ehime Univ.)
Secretary Hidetsugu Irie(Utsunomiya Univ.) / Takashi Miyoshi(Hokkaido Univ.) / Hiroshi Takahashi(Tokyo Inst. of Tech.) / (Nihon Univ.)
Assistant Yasuaki Ito(Hiroshima Univ.) / Tomoaki Tsumura(Nagoya Inst. of Tech.)

Paper Information
Registration To Technical Committee on Computer Systems / Technical Committee on Dependable Computing / Special Interest Group on System Architecture
Language JPN
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) Optimization of Numerical Expression in CNN using Genetic Algorithm
Sub Title (in English)
Keyword(1) CNN
Keyword(2) Adaptive Quantization
Keyword(3) Genetic Algorithm
Keyword(4) Numerical Expression
Keyword(5) Optimization
1st Author's Name Wakana Nogami
1st Author's Affiliation The University of Tokyo/National Institute of Advanced Industrial And Technology(Tokyo Univ./AIST)
2nd Author's Name Tsutomu Ikegami
2nd Author's Affiliation National Institute of Advanced Industrial And Technology(AIST)
3rd Author's Name Shin-ichi O'uchi
3rd Author's Affiliation National Institute of Advanced Industrial And Technology(AIST)
4th Author's Name Ryosei Takano
4th Author's Affiliation National Institute of Advanced Industrial And Technology(AIST)
5th Author's Name Yuma Kishi
5th Author's Affiliation The University of Tokyo/National Institute of Advanced Industrial And Technology(Tokyo Univ./AIST)
6th Author's Name Tomohiro Kudoh
6th Author's Affiliation The University of Tokyo/National Institute of Advanced Industrial And Technology(Tokyo Univ./AIST)
Date 2018-08-01
Paper # CPSY2018-26
Volume (vol) vol.118
Number (no) CPSY-165
Page pp.pp.193-198(CPSY),
#Pages 6
Date of Issue 2018-07-23 (CPSY)