Exploiting Open-World Data for Adaptive Continual Learning

TMLR Paper3870 Authors

Abstract: Continual learning (CL), which involves learning from sequential tasks without forgetting, is mainly explored in supervised learning settings where all data are labeled. However, high-quality labeled data may not be readily available at a large scale due to high labeling costs, making the application of existing CL methods in real-world scenarios challenging. In this paper, we study a more practical facet of CL: open-world continual learning, where the training data comes from the open-world dataset and is partially labeled and non-i.i.d. Building on the insight that task shifts in CL can be viewed as distribution transitions from known classes to novel classes, we propose OpenACL, a method that explicitly leverages novel classes in unlabeled data to enhance continual learning. Specifically, OpenACL considers novel classes within open-world data as potential classes for upcoming tasks and mines the underlying pattern from them to empower the model's adaptability to upcoming tasks. Furthermore, learning from extensive unlabeled data also helps to tackle the issue of catastrophic forgetting. Extensive experiments validate the effectiveness of OpenACL and show the benefit of learning from open-world data.

Submission Length: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: Rahaf Aljundi

Submission Number: 3870