UNIST(총장 이용훈) 인공지능대학원 백승렬 교수팀이 AI가 기존 지식을 유지하면서도 새로운 정보를 학습할 수 있는 ‘SDDGR(Stability Diffusion-based Deep Generative Replay)’ 기술을 개발했다.
'SDDGR', improving AI learning ability and economic efficiency
The core technology of AI is to learn new information while maintaining existing knowledge. Just as people learn new things without forgetting existing experiences, it is important for AI to implement the same function.
UNIST (President Yong-Hoon Lee) announced on the 20th that Professor Seung-Ryeol Baek's research team at the Graduate School of Artificial Intelligence has developed the 'SDDGR (Stability Diffusion-based Deep Generative Replay)' technology that allows AI to learn new information while maintaining existing knowledge.
'SDDGR' technology enables accurate recognition of AI in areas closely related to daily life, such as smart home appliances, robotics, and the medical field. In particular, it is of great help in autonomous vehicles recognizing various objects on the road and driving safely. When applied to a security system, it can accurately detect intruders and immediately send a warning alarm.
The existing 'Class Incremental Learning (CIL)' technology had limitations in recognizing and classifying multiple objects in an image. To solve this, the 'SDDGR' technology appeared. It creates high-quality images and helps you remember what you've learned before. Through the repetitive process, the quality of the images is further improved and existing knowledge can be effectively maintained. It is about learning more accurately by using methods that improve performance even when learning new data.
It is also economically efficient. Since it does not reuse existing data, it can reduce the cost of storing and processing a large amount of data. It is expected to bring great economic benefits to companies.
Professor Baek Seung-ryeol said, “The SDDGR model will be of great help in improving the accuracy of continuous object detection in various industrial fields.”
“SDDGR technology has shown practical effectiveness in various application fields,” said first author researcher Kim Jun-su. “It will be able to contribute to companies developing better artificial intelligence models at less cost and in less time.”
The results of this study are scheduled to be presented on June 21 at the international computer vision conference CVPR 2024, and were conducted with the support of the Ministry of Science and ICT (MSIT), the National Research Foundation of Korea (NRF), the Institute of Information and Communications Technology Planning (IITP), the Korea Institute of Ocean Science and Technology Promotion (KIMST), LG Electronics, and CJ AI Center.