Metalearning Continual Learning Algorithms

arxiv:2312.00276

Original source · Paper JSON

Reinforcement learning · 2023-12-01

Cites (2) · Cited by (4)

Complete cleaned Markdown · Source Markdown including bibliography

Read in parts

1. Part 1
2. Part 2
3. Part 3
4. Part 4
5. Part 5

Contents

• Metalearning Continual Learning Algorithms
• Abstract
• 1 Introduction
• 2 Background
• 2.1 Continual Learning
• 2.2 Metalearning via Sequence Learning a.k.a. In-Context Learning
• 2.3 Self-Referential Weight Matrices and Recursive Self-Transformers
• General description.
• 3 Method
• 4 Experiments
• 4.1 Two-Task Setting: Comprehensible Study
• 4.2 Analysis: Emergence of In-Context Catastrophic Forgetting
• 4.3 General Evaluation
• 5 Discussion
• Other Limitations.
• 6 Conclusion
• Acknowledgements
• Appendix A Experimental Details
• A.1 Continual and Metalearning Terminologies
• Continual learning.
• Metalearning.
• A.2 Datasets
• A.3 Training Details & Hyper-Parameters
• A.4 Evaluation Procedure
• A.5 ACL Objectives with More Tasks
• A.6 Auxiliary 1-shot Learning Objective
• A.7 Details of the Split-MNIST experiment
• A.8 Details of the Split-CIFAR100 and 5-datasets Experiment using ViT
• Appendix B Extra Experimental Results
• B.1 Ablation Studies on the Choice of Meta-Validation Dataset
• B.2 Performance on Split-Omniglot
• B.3 Varying the Number of In-Context Examples/Demonstrations
• B.4 Varying the Number of Tasks in the ACL Meta-Training Objective
• B.5 Further Discussion on Limitations
• Domain generalization.
• Length generalization.
• B.6 More Visualizations
• Instructions for reporting errors