GIST, Hanwha Vision top international acoustic AI competition with 'continual learning' breakthrough

A joint research team from the Gwangju Institute of Science and Technology and Hanwha Vision claimed the top overall team ranking at the DCASE 2026 Challenge, an international acoustic AI competition, GIST announced Tuesday.

The result reflects the team's world-class capability in "continual learning" — an acoustic AI technology that acquires knowledge of sounds from new environments without losing what it has already learned.

The joint team entered Task 7 of the competition's seven tasks, focused on "incremental learning for domain-agnostic audio classification."

The task required competitors to accurately classify 10 types of sounds — including a baby crying, a dog barking and a fire alarm — under "domain-agnostic" conditions, meaning the AI receives no information about the environment in which the audio was recorded.

Conventional audio recognition AI has long struggled with two problems that limit real-world deployment: performance degradation when recording equipment, location or ambient noise changes, and "catastrophic forgetting," in which the model loses previously learned information as it trains on new data.

To address these shortcomings, the team developed continual learning technology that lets the AI retain its existing sound-recognition ability while training on new environments.

The research team preserved each AI model snapshot at every training stage so new learning would not overwrite existing knowledge.

The team also applied a "deep-inversion-based generative replay" technique, which reconstructs the acoustic features of previously learned sounds without repeatedly storing or retraining on the original data — minimizing the loss of prior knowledge during new-environment training.

The team's final submitted system recorded a mean accuracy of 79.62 percent, securing first place in the official team rankings.

The team also swept the official system rankings, with all four submitted systems — three individual models and one ensemble combining them — placing first through fourth.

"This achievement is global recognition of our continual-learning-based acoustic AI technology, which learns new environmental sounds without forgetting what it has already acquired," said Kim Hong-kuk, a GIST professor who led the research. "Combined with Hanwha Vision's intelligent video security technology, it has the potential to evolve into a next-generation AI security solution capable of analyzing video and audio information simultaneously."