Innovation in Autonomous Driving and Security Systems... Development of AI Technology for Maintaining and Learning Existing Knowledge

Professor Baek Seungryul's Team at UNIST Graduate School of Artificial Intelligence Develops 'SDDGR'

Enhanced AI Learning Ability and Economic Efficiency Offer Major Benefits to Companies

The core technology of AI is the ability to learn new information while retaining existing knowledge. Just as humans do not forget past experiences when learning something new, it is important for AI to implement the same function.

The research team led by Professor Baek Seungryul at the Graduate School of Artificial Intelligence at UNIST (President Lee Yonghoon) has developed a technology called 'SDDGR (Stability Diffusion-based Deep Generative Replay)' that enables AI to learn new information while maintaining existing knowledge.

Professor Seungryul Baek.

The 'SDDGR' technology enables precise AI recognition in areas closely related to daily life, such as smart home appliances, robotics, and the medical field. In particular, it is highly beneficial for autonomous vehicles to recognize various objects on the road and operate safely. When applied to security systems, it can accurately detect intruders and immediately send warning alarms.

Previously developed 'Class Incremental Learning (CIL)' technology had limitations in recognizing and classifying multiple objects within an image. To address this, 'SDDGR' was introduced. It generates high-quality images to help AI remember what it has previously learned. Through repeated processes, the quality of the images is further enhanced, allowing for more effective retention of existing knowledge. It also uses methods to improve performance when learning new data, enabling more accurate learning.

It is also highly economically efficient. Since it does not repeatedly use existing data, it can reduce the costs of storing and processing large-scale data. This is expected to bring significant economic benefits to companies.

Professor Baek Seungryul said, "The SDDGR model will be of great help in improving the accuracy of continuous object detection across various industries."

First author Kim Junsu stated, "The SDDGR technology has demonstrated practical effectiveness in various application fields," adding, "It can contribute to companies developing better AI models with less cost and time."

Operation process of the SDDGR (Stable Diffusion-based Deep Generative Reconstruction) model.

The results of this research will be presented at CVPR 2024, a world-renowned computer vision conference, on June 21. The study was supported by the Ministry of Science and ICT (MSIT), the National Research Foundation of Korea (NRF), the Institute of Information & Communications Technology Planning & Evaluation (IITP), the Korea Institute of Marine Science & Technology Promotion (KIMST), LG Electronics, and CJ AI Center.

This content was produced with the assistance of AI translation services.