Work

Continual learning: A systematic literature review

Year: 2025

Type: review

Abstract: In today's data-driven landscape, intelligent systems must continuously learn from evolving data streams while preserving previously acquired knowledge. Continual learning (CL) has emerged as a key paradigm to meet this need, offering the potential to build adaptable and resilient AI systems. Despite its promise, CL faces core challenges such as catastrophic forgetting (CF), which occurs when learning new tasks degrades prior knowledge. These challenges have spurred rapid growth in CL research and method development. This paper presents a systematic literature review (SLR) of CL research published from 2010 to 2025. We adopt a reproducible methodology that combines automated searches across five major databases with manual searches from 13 key venues, resulting in a curated set of 189 peer-reviewed papers. Our review fills key gaps in earlier surveys by unifying fragmented classifications, probing overlooked causes of catastrophic forgetting, and advancing a more principled theoretical framing of CL challenges beyond empirical benchmarking. We contribute by offering (1) an updated, comprehensive overview of CL research; (2) refined taxonomies of learning methods and settings; (3) an analysis of the structural and functional causes of CF; and (4) the identification of open theoretical questions and emerging research directions. We provide an open-source repository of the reviewed literature and associated resources to support ongoing work: https://github.com/yyy24601/awesome-continual-learning.

Source: Neural Networks

DOI: 10.1016/j.neunet.2025.108226

Authors: Qinwen Yang, Liyuan Wang, Joerg Wicker, Gillian Dobbie

Institutions: University of Auckland, Tsinghua University

Language: English

Topic: Domain Adaptation and Few-Shot Learning

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: hybrid

Estimated APC paid: $3,110

Funder: China Scholarship Council