Unsupervised continual learning based on parameter isolation

Abstract

Continual learning (CL) is crucial for the development of intelligent systems that must evolve over time and learn from new experiences without forgetting previously acquired knowledge. Supervised Continual Learning (SCL) addresses this issue by adapting to changing data distributions with labeled data. In real-world applications, such as smart healthcare, robotics, autonomous systems, speech translation, etc., the ability to learn continuously is vital, especially when labeled data are sparse or unavailable. Unsupervised continual learning (UCL) addresses this challenge by enabling models to learn from unlabeled data, avoiding the need for manual annotation. The prominent obstacle in UCL is mitigating catastrophic forgetting (CF), where models forget previously learned knowledge when exposed to new information.