Method for assessing test adequacy of neural network based on element decomposition

US12306745B2 · kind B2 · utility

Assignee

• BEIJING AEROSPACE INSTITUTE FOR METROLOGY AND MEASUREMENT TECHNOLOGY · CN

Inventors

• Yinxiao MIAO · Beijing, CN
• Yifei Liu · Beijing, CN
• Ping-Che YANG
• Xiujian Zhang · Beijing, CN
• Zhonghao Cheng · Beijing, CN
• Long Zhang · Issaquah, US
• Tianqi Wan · Beijing, CN
• Haoyi Chen · Beijing, CN
• Jing Sun · Suzhou, CN
• Yijia Ding · Beijing, CN

Key dates

• Filing date: Nov 22, 2023
• Grant date: May 20, 2025
• Expiry date: Jan 29, 2044

Classification

• Technology area (CPC G): Physics
• CPC primary: G06F11/3688
• WIPO field: Computer technology
• WIPO sector: Electrical engineering

Abstract

A method for assessing test adequacy of deep neural networks based on element decomposition is provided. The network testing is divided into black box testing and white box testing, of which key elements are decomposed and defined. Network parameters including a weight matrix and a bias vector are extracted. Importance values of neurons in individual layers of the deep neural network are calculated and clustered, and an importance value hot map of neurons in each layer is generated based on clustering results. Mutation testing, and index calculation and evaluation are performed.