Title:
NEURAL NETWORK MODEL COMPRESSION METHOD AND NEURAL NETWORK MODEL COMPRESSION DEVICE
Document Type and Number:
WIPO Patent Application WO/2023/171930
Kind Code:
A1
Abstract:
A neural network model compression method according to an embodiment described in the present application comprises the steps of: acquiring execution data of a trained large neural network model and compression ratio information for compressing the large neural network model; examining the structure of the large neural network model on the basis of the execution data and acquiring layer information about the large neural network model; acquiring initial weights by adjusting the weight size of the large neural network model on the basis of the layer information and the compression ratio information, and acquiring a small neural network model having the initial weights; and updating the initial weights of the small neural network model so that the performance of the small neural network model approximates the performance of the large neural network model.
More Like This:
Inventors:
KWON MIN SU (KR)
LEE JOO HO (KR)
LEE JOO HO (KR)
Application Number:
PCT/KR2023/002282
Publication Date:
September 14, 2023
Filing Date:
February 16, 2023
Export Citation:
Assignee:
ENERZAI INC (KR)
International Classes:
G06N3/08; G06N3/04; G06N5/04
Domestic Patent References:
| WO2021072236A2 | 2021-04-15 |
Foreign References:
| KR20210093931A | 2021-07-28 | |||
| US20210255799A1 | 2021-08-19 | |||
| KR20210045274A | 2021-04-26 |
Other References:
CHEON, HEE-SEON ET AL.: "Dynamic Filter Pruning with Decorrelation Regularization for Compression of Deep Neural Network", PROCEEDINGS OF THE KSC 2020, 2020, pages 1313 - 1315, XP009548924
Attorney, Agent or Firm:
DODAM IP LAW FIRM (KR)
Download PDF: