Presentation 2023-10-11
Low power consumption of neural networks using memory bit reduction and sampling
Hirokazu Kodachi, Masato Kitakami,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) Advances in machine learning and the use of big data have increased the computational complexity of neural networks. As a result, the increase in power consumption of neural networks has become a problem. As a solution to this problem, we focused on the structure of neural networks. Neural networks have a structure in which features are distributed and propagated, so some errors can be tolerated. Therefore, there is a method to reduce the power consumption of neural network weights while maintaining recognition accuracy by reducing the mantissa part of the neural network weights that can tolerate bit errors. In this paper, we investigate the changes in power consumption and recognition accuracy by changing the data type of the memory used to store weights and reducing the data to be computed by sampling.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) Neural Network / Low power consumption
Paper # SS2023-26,DC2023-32
Date of Issue 2023-10-04 (SS, DC)

Conference Information
Committee SS / DC
Conference Date 2023/10/11(2days)
Place (in Japanese) (See Japanese page)
Place (in English)
Topics (in Japanese) (See Japanese page)
Topics (in English)
Chair Kozo Okano(Shinshu Univ.) / Tatsuhiro Tsuchiya(Osaka Univ.)
Vice Chair Yoshiki Higo(Osaka Univ.) / Toshinori Hosokawa(Nihon Univ.)
Secretary Yoshiki Higo(Shinshu Univ.) / Toshinori Hosokawa(Tokyo Inst. of Tech.)
Assistant Shinsuke Matsumoto(Osaka Univ.)

Paper Information
Registration To Technical Committee on Software Science / Technical Committee on Dependable Computing
Language JPN
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) Low power consumption of neural networks using memory bit reduction and sampling
Sub Title (in English)
Keyword(1) Neural Network
Keyword(2) Low power consumption
1st Author's Name Hirokazu Kodachi
1st Author's Affiliation Chiba University(Chiba Univ.)
2nd Author's Name Masato Kitakami
2nd Author's Affiliation Chiba University(Chiba Univ.)
Date 2023-10-11
Paper # SS2023-26,DC2023-32
Volume (vol) vol.123
Number (no) SS-206,DC-207
Page pp.pp.27-31(SS), pp.27-31(DC),
#Pages 5
Date of Issue 2023-10-04 (SS, DC)