Title:
ELECTRONIC DEVICE AND CONTROL METHOD THEREFOR
Document Type and Number:
WIPO Patent Application WO/2022/097902
Kind Code:
A1
Abstract:
An electronic device and a control method therefor are disclosed. The electronic device of the present disclosure may comprise: a memory for storing first input data and first weight data used for the operation of a neural network model; and a processor for obtaining second input data obtained by removing a part of the first input data and second weight data obtained by removing a part of the first weight data, by inputting the first input data and the first weight data to a first module, obtaining first output data by inputting the second input data and the second weight data to a second module for performing a multiplication operation, and obtaining second output data by converting the obtained first output data to the form of a floating point indicating 1 bit in a unit scale, on the basis of scaling factors of the first input data and first weight data, the scaling factors being identified through the first module.
Inventors:
YUN JEONGIN (KR)
CHOI SUNGPILL (KR)
LEE JONGHUN (KR)
CHOI SUNGPILL (KR)
LEE JONGHUN (KR)
Application Number:
PCT/KR2021/012448
Publication Date:
May 12, 2022
Filing Date:
September 14, 2021
Export Citation:
Assignee:
SAMSUNG ELECTRONICS CO LTD (KR)
International Classes:
G06N3/08; G06F5/01; G06N3/04
Foreign References:
| CN110780845A | 2020-02-11 | |||
| US20200117981A1 | 2020-04-16 | |||
| CN111178508A | 2020-05-19 | |||
| US20200234126A1 | 2020-07-23 | |||
| KR20190043849A | 2019-04-29 |
Other References:
WANG NAIGANG, CHOI JUNGWOOK, BRAND DANIEL, CHEN CHIA-YU, GOPALAKRISHNAN KAILASH: "Training Deep Neural Networks with 8-bit Floating Point Numbers", NIPS`18: PROCEEDINGS OF THE 32ND INTERNATIONAL CONFERENCE ON NEURAL INFORMATION PROCESSING SYSTEMS (NEURIPS 2018), 3 December 2018 (2018-12-03), XP055927989
Attorney, Agent or Firm:
KIM, Tae-hun et al. (KR)
Download PDF: