A Comprehensive Survey of Continual Learning: Theory, Method and Application
By Liyuan Wang, Xingxing Zhang, Hang Su, Jun Zhu. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023
1,490 citations recorded by the source corpus.
To cope with real-world dynamics, an intelligent system needs to incrementally acquire, update, accumulate, and exploit knowledge throughout its lifetime. This ability, known as continual learning, provides a foundation for AI systems to develop themselves adaptively. In a general sense, continual learning is explicitly limited by catastrophic forgetting, where learning a new task usually results in a dramatic performance drop of the old tasks. Beyond this, increasingly numerous advances have emerged in recent years that largely extend the understanding and application of continual learning. The growing and widespread interest in this direction demonstrates its realistic significance as well as complexity. In this work, we present a comprehensive survey of continual learning, seeking to bridge the basic settings, theoretical foundations, representative methods, and practical applications. Based on existing theoretical and empirical results, we summarize the general objectives of continual learning as ensuring a proper stability-plasticity trade-off and an adequate intra/inter-task generalizability in the context of resource efficiency. Then we provide a state-of-the-art and elaborated taxonomy, extensively analyzing how representative strategies address continual learning, and how they are adapted to particular challenges in various applications. Through an in-depth discussion of promising directions, we believe that such a holistic perspective can greatly facilitate subsequent exploration in this field and beyond.
Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Infectious Diseases, Meta & Continual Learning
View Paper: https://arxiv.org/abs/2302.00487
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Artificial Intelligence
- Jun Zhu
- Hang Su
- Liyuan Wang
- Joost van de Weijer
- Philip H. S. Torr
- Yunchao Wei
- Xingxing Zhang
- Xiaodan Liang
In Computer Vision and Pattern Recognition
- Luc Van Gool
- Dacheng Tao
- Ming-Hsuan Yang
- Yi Yang
- Trevor Darrell
- Xiaogang Wang
- Alan L. Yuille
- Andrea Vedaldi
In Infectious Diseases
In Meta & Continual Learning
- Liyuan Wang
- Xingxing Zhang
- Jun Zhu
- Hang Su
- Joost van de Weijer
- Jiangpeng He
- Bartłomiej Twardowski
- Liyuan Wang
Citations in the corpus
- Incorporating neuro-inspired adaptability for continual learning in artificial intelligence (2023, background): “…three evaluation metrics for visual classification tasks, such as average accuracy (AAC), forward transfer (FWT) and backward transfer (BWT) [6, 75]: where A t,i is the test accuracy of task i after continual learning of task t , and ˜ a i is the test accuracy of each task i learned from…”
- SLCA: Slow Learner with Classifier Alignment for Continual Learning on a Pre-trained Model (2023, background): “On the other hand, since the large amount of pre-training data are typically unlabeled and may also arrive incrementally, it seems more reasonable to use self-supervised pre-training than supervised pre-training [5, 43], also regarding that (upstream) continual learning in a self-supervised manner i”
- Towards Label-Efficient Incremental Learning: A Survey (2023, background): “While there have been few studies that surveys conventional incremental learners [Masana et al., 2020; De Lange et al., 2021; Wang et al., 2023; Zhou et al., 2023], their focus is still on the old label-hungry paradigm in incremental learning.”
- A soft nearest-neighbor framework for continual semi-supervised learning (2022, background): “These challenges result in vanilla CL methods underperforming, as they lack the ability to extract information from the unlabeled set, thus largely overfitting to the labeled set [52].”
- Continual Barlow Twins: Continual Self-Supervised Learning for Remote Sensing Semantic Segmentation (2022, methodology): “Thus, when dealing with multiple datasets, most of the classes can change and we want our method to naturally adapt to this shift by considering separate heads in a task incremental scenario, common in CL [68], [69].”
- Towards Backward-Compatible Continual Learning of Image Compression (2024, background): “It is worth noting that this backward compatibil ity problem is different from the well-known problem of catastrophic forgetting in neural networks [28).”
- Consistent Prompting for Rehearsal-Free Continual Learning (2024, neutral): “The continual learning performance of classification models is mainly evaluated by two metrics [42]: the average accuracy of all classes after learning the last task (denoted as Last-acc) and the averaged incremental accuracy over all learned tasks (denoted as Avg-acc).”
- Orchestrate Latent Expertise: Advancing Online Continual Learning with Multi-Level Supervision and Reverse Self-Distillation (2024, background): “Continual Learning (CL) aims to improve the adaptability of AI systems to the ever-evolving environment of the real world [35, 42, 47, 66, 74].”
- Resurrecting Old Classes with New Data for Exemplar-Free Continual Learning (2024, background): “Recent studies in continual learning (CL) [4, 29, 50, 54] focus on two prevalent scenarios [47]: Task-Incremental Learning (TIL), where task information is available during testing, and Class-Incremental Learning (CIL), where it is not.”
- Continual Learning on a Diet: Learning from Sparsely Labeled Streams Under Constrained Computation (2024, background): “Due to the diversity of the real-world data flow, there has been a lot of work exploring how to eliminate specific constraints in continual learning (Wang et al., 2024), thus making them applicable.”
- InfLoRA: Interference-Free Low-Rank Adaptation for Continual Learning (2024, background): “Through (2), we can find that each column of ∂ L ∂ W can be represented as input h multiplied by a real value a k ( 1 ≤ k ≤ d O ). of Therefore, in the linear layer, each column the gradient ∂ L ∂ W lies in the span of input.”
- Voyager: An Open-Ended Embodied Agent with Large Language Models (2023, background): “[32] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu.”
- CTP: Towards Vision-Language Continual Pretraining via Compatible Momentum Contrast and Topology Preservation (2023, background): “Prior studies on continual learning [28, 80, 65, 78, 76] focused on supervised class-incremental learning (CIL), as Figure 1(a), which aims to maintain discriminative features for known classes and expand new classifiers to learn new classes.”
- Improving Plasticity in Online Continual Learning via Collaborative Learning (2023, background): “CL can be done in two different manners [4, 47]: offline continual learning and online continual learning.”
- UniSD: Towards a Unified Self-Distillation Framework for Large Language Models (2026, unknown): “Continual learning [37] aims to adapt models to new knowledge and skills while preserving existing capabilities, a challenge known as catastrophic forgetting [36, 38].”