The paper introduces a new paradigm in Continual Learning named Realistic Continual Learning (RealCL), where class distributions are randomized across tasks, distinguishing it from conventional class-incremental learning. The concept of catastrophic forgetting, where models forget previous tasks upon learning new ones, is a central focus. The authors present CLARE, a pre-trained model-based solution to mitigate forgetting and effectively integrate new learning while preserving past knowledge. Through extensive experimentation, CLARE has shown to outperform existing approaches in RealCL scenarios, solidifying its adaptability and robustness. The introduction of RealCL and the development of CLARE are key contributions, offering a more practical and flexible continual learning framework. Key insights from the paper are as follows:

• Introduction of RealCL to generalize classic Continual Learning setups.
• Development of CLARE to merge new knowledge with existing learning.
• Demonstrated superiority of CLARE in various RealCL scenarios.

Why is This Important? This advancement is crucial for developing AI that can learn continuously and adapt in dynamic environments without losing prior knowledge. It echoes the human ability to learn throughout life, marking a significant step toward more natural and efficient learning processes in AI. Read more…