Addressing catastrophic forgetting in class-incremental learning—a survey

Abstract Class-Incremental Learning (CIL) enables models to learn new classes over time without forgetting previously acquired knowledge—a process often hindered by Catastrophic Forgetting (CF) . This paper provides a comprehensive and structured overview of the CIL landscape, beginning with a detailed examination of its core variants—Task-, Domain-, and Class-Incremental Learning—and explaining how CIL fits within the broader structure of incremental learning by clarifying the relationships among its major variants and the shared challenges they address. This perspective establishes a consistent conceptual foundation for understanding how different forms of incremental learning relate to continual learning objectives more broadly. The paper then reviews CF in depth, including formal definitions, underlying mechanisms, and evaluation protocols. A central contribution is a refined taxonomy of CIL methods, encompassing replay-based, regularization-based, parameter-isolation, hybrid, and large language model (LLM)-based approaches. Each category is analyzed in terms of method variations, applications, trade-offs, and emerging trends. Quantitative and qualitative comparisons highlight differences in scalability, interpretability, and adaptability to various domains. We also identify open challenges, including bias toward recent classes, memory and compute constraints, and the need for stronger theoretical cohesion in continual adaptation. Looking forward, we outline grounded future directions such as advances in self-supervised and generative replay mechanisms, attention-based transformers, and neuro-symbolic integrations. These directions reflect promising long-term avenues for rethinking and improving representational stability, efficiency, and generalization in class-incremental learning.