Title:
METHOD FOR OPTIMIZING NEURAL NETWORK PARAMETER APPROPRIATE FOR HARDWARE IMPLEMENTATION, NEURAL NETWORK OPERATION METHOD, AND APPARATUS THEREFOR
Document Type and Number:
WIPO Patent Application WO/2020/159016
Kind Code:
A1
Abstract:
The present invention relates to a method for optimizing a neural network parameter appropriate for hardware implementation, a neural network operation method, and an apparatus therefor. The method for optimizing a neural network parameter appropriate for hardware implementation according to the present invention may comprise the steps of: performing type conversion of an existing parameter of a neural network into a size parameter having a single value per channel and a code parameter; and branching out the type-converted size parameter to generate an optimized parameter. Accordingly, the present invention can provide a neural network parameter optimization method, a neural network calculation method, and an apparatus therefor, wherein large operational quantities and parameters which a convolution neural network has are effectively optimized for hardware implementation so that a minimum loss in accuracy and a maximum operational speed can be obtained.
Inventors:
LEE SANG HUN (KR)
KIM MYUNG KYUM (KR)
KIM JOO HYUK (KR)
KIM MYUNG KYUM (KR)
KIM JOO HYUK (KR)
Application Number:
PCT/KR2019/008913
Publication Date:
August 06, 2020
Filing Date:
July 18, 2019
Export Citation:
Assignee:
DEEPER-I CO INC (KR)
International Classes:
G06N3/04; G06N3/063
Domestic Patent References:
| WO2017189859A1 | 2017-11-02 |
Foreign References:
| KR20160143505A | 2016-12-14 | |||
| US20180046900A1 | 2018-02-15 | |||
| US20180181867A1 | 2018-06-28 |
Other References:
GE. SHIMING ET ET AL.: "Compressing deep neural networks for efficient visual inference", PROCEEDINGS OF THE IEEE INTERNATIONAL CONFERENCE ON MULTIMEDIA AND EXPO (ICME) 2017, 14 July 2017 (2017-07-14), pages 667 - 672, XP033146637
See also references of EP 3779801A4
See also references of EP 3779801A4
Attorney, Agent or Firm:
KIM, Bong Jo (KR)
Download PDF: