Continual Meta-Learning refers to a learning paradigm that combines the principles of meta-learning with those of continual learning to develop models that can continuously adapt to new tasks without losing prior knowledge.
Formally, a model \( M_\theta \) learns a sequence of tasks \( \{T_1, T_2, ..., T_n\} \), where for each task a rapid adaptation with few examples is enabled while simultaneously minimizing catastrophic forgetting.
This is often realized through optimization strategies that integrate both meta-learning updates and memory or regularization mechanisms for previous tasks.
Definition:
“Continual Meta-Learning is a learning paradigm that enables a model to continuously learn new tasks and quickly adapt to them while preserving previously acquired knowledge and preventing catastrophic forgetting.”
Source:
Finn, C., Rajeswaran, A., Kakade, S., & Levine, S. (2020). Online Meta-Learning. Proceedings of the 8th International Conference on Learning Representations (ICLR).