First Place in DKIS 2026 Challenge Category
Offering Solutions to Difficulties with "Continual Learning Technology"
Accelerating the Development of Next-Generation AI Security Solutions

Hanwha Vision and a research team from the Gwangju Institute of Science and Technology (GIST) announced on July 14 that they secured first place at the international acoustic artificial intelligence (AI) competition, "DKIS 2026 Challenge," with their jointly developed "Continual Learning" technology.


This achievement was made possible through collaboration between the AI Research Center at Hanwha Vision R&D Center and the team led by Professor Kim Hongguk from the Department of Electrical, Electronic, and Computer Engineering at GIST. Their work addresses a critical weakness of AI: the tendency to forget previously learned sounds when learning new ones.

Screenshot of the awards results page on the 'DKIS 2026' website

Screenshot of the awards results page on the 'DKIS 2026' website

View original image

The "DKIS Challenge," hosted by the Institute of Electrical and Electronics Engineers (IEEE), is an international competition where academic and corporate research teams compete in acoustic AI technology. This year, a total of 135 teams participated.


The joint research team ranked first among 21 teams in the "Incremental Learning for Domain-Agnostic Audio Classification" category, which is one of seven divisions. They successfully completed a mission to accurately distinguish 10 types of sounds—such as baby cries, dog barking, and fire alarms—even in environments where the source of the sound could not be identified.


The team focused particularly on solving the problem of "catastrophic forgetting." Traditional audio recognition AI systems suffer performance degradation when the environment changes, such as differences in recording devices, locations, or background noise. More critically, when learning new data, the AI tends to forget previously learned information, which becomes a major obstacle to accumulating data. For example, an AI that previously excelled at distinguishing urban noise may struggle to recognize city sounds again after learning acoustic data from an airport environment.


This problem arises because AI models designed for specific purposes are generally smaller in scale compared to ultra-large, general-purpose AI models. While the phenomenon is less frequent in large-scale AI models, it fundamentally occurs there as well. In fields such as security and industrial safety, where external environments change constantly, this issue is recognized as a core challenge for the advancement of AI technology.


The continual learning technology developed by the joint research team incorporates a "deep inversion-based generative replay" technique, which leverages and reproduces previously learned sound features. Deep inversion is a technology that enables the classification model for each sound to autonomously regenerate and reverse-engineer acoustic data similar to what it learned in the past, without additional data collection. For instance, even if an AI installed in an indoor monitoring system learns new road noise, it can still trace back the acoustic features—such as decibel levels and amplitudes—of previously learned indoor sounds, thus maintaining its classification ability. With this approach, the team successfully minimized the loss of previously acquired knowledge during new learning stages.


It is also notable that the team achieved high accuracy and stable performance by synthesizing the prediction results from multiple AI models using an "ensemble technique." The final submitted system recorded the highest average accuracy score of 79.62% among all competitors. The average accuracy of the remaining 20 teams was approximately 68%, making the joint research team's result about 10% higher. Furthermore, the final ensemble system and three additional single systems submitted by the joint team ranked first through fourth in the official system rankings for the category, demonstrating their technical prowess.


This technology, which allows for the stable accumulation of acoustic information in various environments, is expected to be actively utilized in the development of Hanwha Vision's next-generation AI security solutions.


Jeongeun Lim, Head of the AI Research Center at Hanwha Vision, stated, "This award proves that the direction and technical level of Hanwha Vision's research are world-class. The 'continual learning' technology is directly linked to security camera adaptability in external environments, and we will continue to develop this technology as a future solution."


The "DKIS 2026 Workshop," where the research results will be presented and the winning team will be awarded, is scheduled to take place in Boston, USA, for two days starting October 28.



This research, conducted through industry-academia collaboration, was supported by Hanwha Vision's own research project as well as funding from the Ministry of Trade, Industry and Energy.


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

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing