Research presents new framework in online video perception

Kim Min-woo, a senior in Kookmin University’s department of software engineering / Courtesy of Kookmin University
 

A research paper on artificial intelligence (AI) co-authored by a Kookmin University student has been accepted for presentation at the 43rd International Conference on Machine Learning (ICML), the school said Monday.

Kim Min-woo, a senior in the university’s department of software engineering, is the first author of the paper titled “Memory as Dynamics: Learning Reliability-Guided Predictive Models for Online Video Perception.”

The university said the paper presents a new framework in the field of online video perception that reinterprets video memory as a “dynamic latent process” rather than a “static buffer.”

In the paper, Kim introduced reliability-guided predictive memory (RPM), a framework that explicitly regulates when and how predictive dynamics should influence online video perception.

His research demonstrated that the new framework utilized temporal information in video sequences more accurately and efficiently.

By dynamically estimating the reliability of each video frame and incorporating this assessment into memory updates and predictions, the study showed that stable recognition performance can be maintained even in the presence of noise or occlusion.

Experimental results demonstrated that the proposed method outperformed existing approaches across various online video benchmarks. The research is significant in that it could improve both the accuracy and stability of video recognition systems.

The university expects the proposed framework to be applied in various fields such as autonomous driving, robotics and video-understanding systems where real-time video recognition is critical.

“I sought to reinterpret the relationship between memory and prediction from a new perspective,” Kim said. “It is very meaningful that the research I conducted as an undergraduate has been recognized at a world-class academic conference.”

He added, “I hope to continue practical AI research that can contribute to solving real-world problems.”

His research was supported by the National Research Foundation of Korea and the Institute of Information and Communications Technology Planning and Evaluation.

ICML 2026 is scheduled to be held in Seoul from July 6-11. ICML is widely regarded as one of the world’s most prestigious conferences in machine learning and AI, alongside the Conference on Neural Information Processing Systems (NeurIPS) and International Conference on Learning Representations (ICLR).

Since its inception in 1980, the conference has established itself as a key platform where researchers from across the globe present the latest machine learning research findings every year.