A research team led by Prof. Kim Jang Ho of the Department of Artificial Intelligence at KU, including students Shin Hyun Yun(MSc) and Han Seung Jin, presented the paper “Cooperative Meta-Learning with Gradient Augmentation” at the UAI conference held at Universitat Pompeu Fabra in Spain on Tuesday, July 16th. The UAI conference is the premier conference on artificial intelligence and has been held annually since 1985, celebrating its 40th edition this year.

In this paper, we developed Cooperative Meta-Learning (CML) to improve the performance of meta-learning and proposed a new method to increase the generalization performance of models by extending the existing Model-Agnostic Meta-Learning (MAML) structure.

While traditional MAML learns new tasks through two optimization loops, CML introduces a co-learner to help it better find meta-initialization parameters. The co-learner is trained only in the outer loop without updating the inner loop, providing a different perspective in the process of finding meta-initialization parameters. This gives the model the ability to quickly adapt to new tasks. Through several experiments, the researchers demonstrated that CML outperforms existing meta-learning methods, especially in the problem of few-shot learning, which requires learning a new task with a small amount of data.

Prof. Kim Jang Ho's research team has released the code of CML so that it can be easily applied to various meta-learning studies, and plans to continue innovative research in the future.