元学习——MAML、Reptile与ANIL
作者:凯鲁嘎吉 - 博客园 http://www.cnblogs.com/kailugaji/
之前介绍过元学习——从MAML到MAML++,这次在此基础上进一步探讨,深入了解MAML的本质,引出MAML高效学习的原因究竟是快速学习,学到一个很厉害的初始化参数,还是特征重用,初始化参数与最终结果很接近?因此得到ANIL(Almost No Inner Loop),随后我们阅读了Reptile——On first-order meta-learning algorithms,另一种元学习方法,并比较了MAML、Reptile与模型预训练之间的区别。
1. Meta Learning vs Machine Learning
2. MAML vs Model Pre-Training
3. MAML——Feature Reuse
4. MAML vs ANIL
MAML的目标是学习一个参数𝜃使得其经过一个梯度迭代就可以在新任务上达到最好的性能。
内循环:与具体任务相关的任务适配参数的更新(自适应具体任务)
外循环:整体任务空间上的模型参数的更新(元初始化)
5. Reptile: On First-Order Meta-Learning Algorithms
6. MAML, Model Pre-Training, and Reptile
7. 参考文献
[1] GitHub - Fafa-DL/Lhy_Machine_Learning: 李宏毅2021春季机器学习课程课件及作业 https://github.com/Fafa-DL/Lhy_Machine_Learning
[2] Finn, C., Abbeel, P. & Levine, S. Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks. ICML 2017.
[3] Aniruddh Raghu, Maithra Raghu, Samy Bengio, Oriol Vinyals, Rapid Learning or Feature Reuse? Towards Understanding the Effectiveness of MAML, ICLR, 2020.
[4] Nichol A, Achiam J, Schulman J. On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999, 2018.
[5] Nichol A, Schulman J. Reptile: a scalable metalearning algorithm. arXiv preprint arXiv:1803.02999, 2018, 2(3): 4.
[6] Reptile: A Scalable Meta-Learning Algorithm. OpenAI. https://openai.com/blog/reptile/
[7] 一文入门元学习(Meta-Learning)(附代码) - 知乎 https://zhuanlan.zhihu.com/p/136975128