Title:
SAMPLE PRINCIPAL COMPONENT ANALYSIS-BASED ARCHITECTURE SEARCH METHOD IN IMAGE CLASSIFICATION
Document Type and Number:
WIPO Patent Application WO/2024/045375
Kind Code:
A1
Abstract:
A sample principal component analysis-based architecture search method in image classification. On the basis of DARTS, the method brings forward the importance of calculating an op in a classification network on the basis of sample principal component analysis. Compared with an original DARTS which directly uses weight parameters to select a classification network op, the method provided can find a classification network architecture having better performance. In previous architecture search methods, the importance of an op is usually evaluated mainly according to the sizes of weight parameter values after training is finished, so as to build a final classification network, and said methods select only by referring to the weight parameter values after the training is finished, thus causing a relatively large error. However, the present method considers all weight parameter values during a model training process, performs sample principal component analysis on the weight parameter values, and selects an important OP to build a classification network.
Inventors:
LI HUI (CN)
FANG XUWEI (CN)
XU XIAOLONG (CN)
ZHOU SONG (CN)
FANG XUWEI (CN)
XU XIAOLONG (CN)
ZHOU SONG (CN)
Application Number:
PCT/CN2022/134583
Publication Date:
March 07, 2024
Filing Date:
November 28, 2022
Export Citation:
Assignee:
CHINA TELECOM BESTPAY CO LTD (CN)
International Classes:
G16H30/20
Foreign References:
| CN111931904A | 2020-11-13 | |||
| CN102915445A | 2013-02-06 | |||
| CN108829816A | 2018-11-16 | |||
| CN109376787A | 2019-02-22 | |||
| US20220035878A1 | 2022-02-03 |
Download PDF: