Reptile: On First-Order Meta-Learning Algorithms

On First-Order Meta-Learning Algorithms

Paper:https://arxiv.org/pdf/1803.02999.pdf
Code:https://github.com/openai/supervised-reptile
Tips：OpenAi的一篇相似MAML的Meta-learning相关的paper。
（阅读笔记）

1.Main idea

• 目标旨在实现相同分布的一类任务的少量样本快速学习。This paper considers meta-learning problems, where there is a distribution of tasks, and we would like to obtain an agent that performs well (i.e., learns quickly) when presented with a previously unseen task sampled from this distribution.
• 类似于first-order MAML，忽略了二阶偏微分。并且指出了其实现更为简单。
• Reptile的作为meta-learning的方法，训练还是和传统方法很相似。Reptile is so similar to joint training that it is especially surprising that it works as a meta-learning algorithm.
• 做出了first-order MAML和reptile的理论分析。

2.MAML回顾

回顾了MAML相关工作。
目标是求解下式，其中$\tau$是不同的任务集，$\varphi$是初始参数，$L$是损失函数，$U_{\tau }^k$表示从任务集$\tau$抽样出来训练的第$k$次的参数更新操作：
$$\underset{\varphi }{\min }{\mathbb{E}}_{\tau }[L_{\tau }(U_{\tau }^k(\varphi ))]$$
$A$是原始训练任务集，$B$是新任务集。MAML的训练操作仍然对原始任务集进行训练，但是其损失函数却是针对的$B$，如下所示：
$$\underset{\varphi }{\min }{\mathbb{E}}_{\tau }[L_{\tau ,B}(U_{\tau ,A}(\varphi ))]$$
找梯度即需对参数$\varphi$求偏导（复合函数求导）：
$$\begin{gathered} g=\frac{\partial L_{\tau ,B}(U_{\tau ,A}(\varphi ))}{\partial \varphi } \\ =L_{\tau ,B}^{\prime }(U_{\tau ,A}(\varphi ))\times U_{\tau ,A}^{\prime }(\varphi )=\frac{\partial L_{\tau ,B}(U_{\tau ,A}(\varphi ))}{\partial U_{\tau ,A}(\varphi )}\times \frac{\partial U_{\tau ,A}(\varphi )}{\partial \varphi } \end{gathered}$$
使用恒等操作（对第二项偏微分变为常量1），得到First-order MAML为：
$$g=\frac{\partial L_{\tau ,B}(U_{\tau ,A}(\varphi ))}{\partial U_{\tau ,A}(\varphi )}$$
即损失下降梯度的方向为在任务集$A$得到参数$\varphi$的情况下，通过对测试集$B$得到的损失最小化的方向即是外循环的方向。

3.Reptile

• 算法流程如下所示：
  
  注意到可以一次迭代中将$\tilde{\varphi }$进行$k$步后，最后才确定梯度的方向。