As far as I understand, episodic learning involves training the model on episodes of tasks, where each task consists of a query set Q and support set S. On each episode the model is trained on the samples in S and then tested on query in Q. How does Meta-learning relate to this? Is episodic learning a precondition for meta-learning or can there be meta-learning without episodic learning or vice versa? All papers that I come across in few-shot learning (in the NER context) seem to use these terms somewhat interchangeably and I never found a real definition. I understand that Meta Learning is a "learning to learn" approach, where the model is updated based on each episode with the hope that it learns to generalize to unseen tasks better. But that must be the case in every episodic learning approach? otherwise it would not make sense to train on episodes.
Asked
Active
Viewed 72 times
