Meta-Learning in Neural Networks: A Survey
Citation: 236 (2021-08-29)
1. Proposed Taxonomy
meta_learning_survey_overview.jpg
1.1 Meta-Representation (What?)
Parameter Initialization
参数初始化最典型的例子是MAML，学习到的meta knowledege是任务的初始化参数，基于这个初始化参数，新的任务可以更快地学习。
参数初始化最大的问题是，外层任务与内层任务可能要学习相同数量的参数。
Optimizer
Feed-Forward Models
Embedding Functions(Metric Learning)
Losses and Auxilliary Tasks
1.2 Meta-Optimizer (How?)
1.3 Meta-Objective (Why?)
References
[1] Hospedales, Timothy, et al. "Meta-learning in neural networks: A survey." arXiv preprint arXiv:2004.05439 (2020).
