Comprehensive survey of continual learning

Fen Lyu, Liang Wang, Li Xi, Wei‐Shi Zheng, Zhang Zhang, Huimin Xiao, Fuyuan HuJournal of Image and Graphics 2025 · DOI

Abstract

Continual Learning (CL) is a pivotal challenge in machine learning, aiming to enable models to learn new tasks sequentially without suffering from "catastrophic forgetting." This survey explores the evolution of CL methodologies, categorizing them into traditional "train-from-scratch" approaches and modern strategies leveraging Pre-trained Models (PTMs). By balancing the "Stability-Plasticity" dilemma, researchers have moved from memory replay and regularization toward lightweight fine-tuning and prompt-based adaptation. Current evidence suggests that PTM-based methods offer superior performance and robustness, defining the current state-of-the-art in the field.

Introduction: The Stability-Plasticity Dilemma

Standard machine learning models typically operate under the assumption of static data distributions. When these models encounter a stream of non-stationary tasks, they often experience Catastrophic Forgetting (CF)—a phenomenon where new information overwrites the weights necessary for previous tasks.

The fundamental challenge in CL is the Stability-Plasticity Dilemma. A model must maintain enough plasticity to acquire new knowledge while ensuring sufficient stability to preserve existing memories. This balance is critical for applications in dynamic environments, such as autonomous driving, robotics, and evolving medical diagnostic systems.

Overview of Continual Learning frameworks, illustrating the Stability-Plasticity dilemma and the transition from traditional methods to PTM-based approaches.

Methodological Evolution

The landscape of CL research can be broadly divided into two eras: traditional continual training (dominant from 2017 to 2024) and the recent shift toward Pre-trained Model-based methods (emerging strongly since 2022).

1. Traditional Continual Training

Traditional methods focus on training models from scratch or from simple initializations. They generally fall into three categories:

  • Memory Replay: These methods store a subset of previous data (exemplars) or use generative models (e.g., GANs) to synthesize old data, which is then interleaved with new task data during training to "rehearse" old knowledge.
  • Regularization: These approaches add constraints to the loss function to protect important parameters. Weight regularization (e.g., EWC) penalizes changes to weights critical for past tasks, while functional regularization uses techniques like knowledge distillation to keep the model's output behavior consistent.
  • Dynamic Architectures: Instead of keeping the model fixed, these methods adapt the network structure. This might involve adding new neurons for new tasks or using sub-network masking to isolate task-specific pathways.

2. PTM-based Continual Learning

The advent of large-scale Pre-trained Models (PTMs) like ViT and BERT has revolutionized CL. PTMs provide a robust foundation of general features, allowing the model to focus on task adaptation rather than learning representations from scratch.

  • Fine-tuning based: Rather than full-parameter updates, these methods use lightweight modules like Adapters or LoRA. By keeping the backbone frozen, the model preserves its general knowledge while adapting to specific task nuances.
  • Prompt-based: This paradigm learns a "Prompt Pool" or specific task prompts (e.g., L2P, DualPrompt). The model learns to select or generate the correct prompt to guide the frozen PTM's features toward the current task, requiring no storage of raw data and minimal parameter overhead.

Performance and Benchmarking

The survey highlights a significant performance gap between traditional and PTM-based methods across standard benchmarks like Split CIFAR-100 and ImageNet-R.

Method Category Benchmark Accuracy Forgetting Rate
Traditional (iCaRL) CIFAR-100 ~66% High
PTM-based (EASE) CIFAR-100 88.11% Low (<6%)
Traditional (EWC) ImageNet-R ~35% High
PTM-based (CPrompt) ImageNet-R 75.30% Low

PTM-based methods consistently demonstrate higher accuracy and significantly lower forgetting rates (often under 6%, compared to 30-50% for some traditional regularization methods).

Future Directions and Challenges

Despite recent progress, several frontiers remain:

  1. Efficiency: PTMs are computationally expensive. Future research must focus on making these methods viable for resource-constrained edge devices.
  2. Unsupervised Learning: Moving away from labeled task streams toward self-supervised continual learning is essential for real-world scalability.
  3. Task Diversity: Current methods struggle when sequential tasks have large domain gaps (e.g., natural images followed by medical scans).
  4. LLM Integration: Leveraging the reasoning capabilities of Large Language Models (LLMs) to enhance visual CL is a promising new avenue.

Conclusion

The field of Continual Learning has undergone a paradigm shift. While traditional methods laid the groundwork for managing the stability-plasticity trade-off, PTM-based approaches—particularly those utilizing prompting—currently offer the most effective solution for mitigating catastrophic forgetting in complex, large-scale scenarios.

Summary updated

Abstract

持续学习(continual learning, CL)是机器学习领域的一个关键问题,旨在使模型在不断学习新任务的同时,避免灾难性遗忘,保持对先前任务的记忆。持续学习已在多个实际应用中扮演重要角色,如自动驾驶、机器人控制和医疗诊断系统等。本文旨在为学界提供持续学习领域的最新研究进展综述,并对未来可能的研究方向进行展望。为实现持续学习中新旧知识学习的“可塑性—稳定性”平衡,国内外研究者们提出多种方法,根据方法的发展路径可以分为传统持续训练方法和基于预训练模型的方法。首先,本文介绍了传统持续训练的关键技术和方法,包括记忆重放法、正则化法和动态结构法。记忆重放法通过将先前任务的样本存储并重放,以帮助模型回忆过去的知识。正则化法则通过对模型参数的更新进行约束,防止新任务对旧任务的干扰。动态结构法通过调整模型结构或引入新的模型模块以应对新任务的挑战,避免灾难性遗忘的发生。接着,本文进一步探讨了基于预训练模型的持续学习方法的进展。随着大规模预训练模型的广泛应用,这类预训练模型展示了强大的泛化能力和知识迁移能力。基于预训练模型的持续学习方法可以分为基于微调和基于提示的方法。微调方法可以通过冻结部分预训练模型参数,仅对特定层进行更新,或采用学习率调节等技术,避免对预训练模型的过度修改。基于提示的方法通过设计和输入提示来引导模型处理新任务,而无需大规模调整模型参数。本文提供的实验结果建议,当前持续学习任务应优先考虑采用基于预训练模型的方法。最后,本文对当前持续学习领域的挑战与未来发展方向进行了展望,重点讨论了各种实际约束条件下,如何结合预训练模型和经典持续学习方法,构建新的架构设计和优化策略,以应对日益复杂的现实任务需求。
BibTeX
@article{lyu2025comprehensive,
  author = {Fen Lyu and Liang Wang and Li Xi and Wei‐Shi Zheng and Zhang Zhang and Huimin Xiao and Fuyuan Hu},
  title = {{Comprehensive survey of continual learning}},
  journal = {Journal of Image and Graphics},
  year = {2025},
  doi = {10.11834/jig.240661},
}
Figure