Three scenarios for continual learning
By Gido M. van de Ven, Andreas S. Tolias. 2019
1,067 citations recorded by the source corpus.
Standard artificial neural networks suffer from the well-known issue of catastrophic forgetting, making continual or lifelong learning difficult for machine learning. In recent years, numerous methods have been proposed for continual learning, but due to differences in evaluation protocols it is difficult to directly compare their performance. To enable more structured comparisons, we describe three continual learning scenarios based on whether at test time task identity is provided and--in case it is not--whether it must be inferred. Any sequence of well-defined tasks can be performed according to each scenario. Using the split and permuted MNIST task protocols, for each scenario we carry out an extensive comparison of recently proposed continual learning methods. We demonstrate substantial differences between the three scenarios in terms of difficulty and in terms of how efficient different methods are. In particular, when task identity must be inferred (i.e., class incremental learning), we find that regularization-based approaches (e.g., elastic weight consolidation) fail and that replaying representations of previous experiences seems required for solving this scenario.
Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Natural Language Processing, Meta & Continual Learning
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
In Computer Vision and Pattern Recognition
- Luc Van Gool
- Dacheng Tao
- Trevor Darrell
- Bernt Schiele
- Ming-Hsuan Yang
- Andrea Vedaldi
- Xiaogang Wang
- Andrew Zisserman
In Natural Language Processing
- Mohit Bansal
- Jianfeng Gao
- Kyunghyun Cho
- Yejin Choi
- Graham Neubig
- Xiang Ren
- Luke S. Zettlemoyer
- Bing Liu
In Meta & Continual Learning
- Bing Liu
- Joost van de Weijer
- Zixuan Ke
- Tinne Tuytelaars
- Liyuan Wang
- Gyuhak Kim
- Jorge Armando Mendez Mendez
- Eric Eaton
Citations in the corpus
- The Curse of Recursion: Training on Generated Data Makes Models Forget (2023, background): “Unlike traditional machine learning which seeks to learn from a static data distribution, continual learning attempts to learn from a dynamic one, where data are supplied in a sequential fashion [Van de Ven and Tolias, 2019].”
- A Comprehensive Survey of Continual Learning: Theory, Method and Application (2023, background): “Such contents could be new skills, new examples of old skills, different environments, different contexts, etc., with particular realistic challenges incorporated [2], [4].”
- Learnability and Algorithm for Continual Learning (2023, methodology): “Two continual learning settings have been popularly studied: task incremental learning (TIL) (van de Ven & Tolias, 2019) and class incremental learning (CIL).”
- SLCA: Slow Learner with Classifier Alignment for Continual Learning on a Pre-trained Model (2023, methodology): “Following previous works of CLPM in CV [45, 44], we mainly focus on the class-incremental scenario of continual learning [36].”
- Towards Label-Efficient Incremental Learning: A Survey (2023, background): “There are three main settings, namely task, domain-and class-incremental learning, see [Van de Ven and Tolias, 2019].”
- FeTrIL: Feature Translation for Exemplar-Free Class-Incremental Learning (2022, background): “Unlike task IL [40], the boundaries be-tween different states are not known at test time.”
- CODA-Prompt: COntinual Decomposed Attention-Based Prompting for Rehearsal-Free Continual Learning (2022, neutral): “[24, 61] parameter counts, but importantly can scale performance by increasing capacity, unlike prior methods which quickly plateau.”
- A Theoretical Study on Solving Continual Learning (2022, background): “There are two main CL settings that have been extensively studied: class incremental learning (CIL) and task incremental learning (TIL) [2].”
- Long-Tailed Class Incremental Learning (2022, background): “Class incremental learning (CIL) is one of the primary scenarios for continual learning [38].”
- NISPA: Neuro-Inspired Stability-Plasticity Adaptation for Continual Learning in Sparse Networks (2022, methodology): “We have mostly evaluated NISPA on task incremental learning (van de Ven & Tolias, 2019) (i.e., task labels are available during testing).”
- Continual Learning with Guarantees via Weight Interval Constraints (2022, methodology): “To verify the empirical usefulness of our method, we test it in three standard continual learning scenarios (Hsu et al., 2018; van de Ven & Tolias, 2019): incremental task, incremental domain, and incremental class.”
- Overcoming Catastrophic Forgetting in Incremental Object Detection via Elastic Response Distillation (2022, background): “Based on whether the task identity is provided or must be inferred [34], researchers divide Incremental Learning (IL) into three types: task/domain/class IL.”
- Class-Incremental Learning by Knowledge Distillation with Adaptive Feature Consolidation (2022, neutral): “However, as discussed in [16, 46], parameter regularization methods empirically present relatively lower accuracy than the techniques in other categories, and might serve as a poor proxy for keeping the network output [8].”
- Self-Sustaining Representation Expansion for Non-Exemplar Class-Incremental Learning (2022, background): “CIL [32, 45] is the most difficult scenery in continual learning and has received more attention recently.”
- On Generalizing Beyond Domains in Cross-Domain Continual Learning (2022, background): “On the other hand, there has been a surge of efforts to devise machine learning based algorithms to build more intelligent models and mitigate the aforementioned challenges from two perspectives, namely continual learning [2, 4, 26, 32] and domain generalSequence of Domains Sequence of Tasks”
- Memory Replay with Data Compression for Continual Learning (2022, neutral): “This setting is also called class-incremental learning (van de Ven & Tolias, 2019).”
- Learning Fast, Learning Slow: A General Continual Learning Method based on Complementary Learning System (2022, background): “There are a plethora of evaluation protocols in the CL literature, each of which biases the evaluation towards a certain approach (Farquhar & Gal, 2018; Mi et al., 2020; Shim et al., 2020; van de Ven & Tolias, 2019).”
- Continual Learning of a Mixed Sequence of Similar and Dissimilar Tasks (2021, background): “Note that in task continual learning (TCL), task ID is needed because in testing each test instance will be tested by the classifier/model of its corresponding task (van de Ven and Tolias, 2019).”