Title:
ELECTRONIC DEVICE AND CONTROL METHOD OF ELECTRONIC DEVICE
Document Type and Number:
WIPO Patent Application WO/2022/270815
Kind Code:
A1
Abstract:
An electronic device and a control method of an electronic device are disclosed. An electronic device according to the present disclosure comprises: a memory for storing first model information including first weight data represented in a real number within a first bit range and information on multiple layers of a neural network model; and a processor for performing recursive quantization including quantization for the first weight data and quantization for error data. In addition, the processor: identifies the number of times quantization is repeated for error data for each of the multiple layers on the basis of the first model information; performs the recursive quantization according to the identified number of times of repetition to acquire quantization information including second weight data corresponding to the first weight data and represented in an integer within the range of a second bit smaller than a first bit; and reconfigures a neural network model on the basis of the first model information, information on the identified number of times of repetition, and information on the second weight data, and acquires second model information for the reconfigured neural network model.
Inventors:
LEE SANGJEONG (KR)
OH JIHUN (KR)
PARK MEEJEONG (KR)
OH JIHUN (KR)
PARK MEEJEONG (KR)
Application Number:
PCT/KR2022/008456
Publication Date:
December 29, 2022
Filing Date:
June 15, 2022
Export Citation:
Assignee:
SAMSUNG ELECTRONICS CO LTD (KR)
International Classes:
G06N3/04; G06N3/08
Foreign References:
| US20170308789A1 | 2017-10-26 | |||
| US20200394523A1 | 2020-12-17 | |||
| US20210193120A1 | 2021-06-24 | |||
| JP2021043905A | 2021-03-18 | |||
| KR20210035702A | 2021-04-01 |
Other References:
SANGHYUN SEO, KIM JUNTAE: "Efficient Weights Quantization of Convolutional Neural Networks Using Kernel Density Estimation based Non-uniform Quantizer", APPLIED SCIENCES, vol. 9, no. 12, pages 2559, XP055699528, DOI: 10.3390/app9122559
Attorney, Agent or Firm:
KIM, Tae-hun et al. (KR)
Download PDF: