# 论文笔记 – Model-agnostic meta-learning for fast adaptation of deep networks

Finn, Chelsea, Pieter Abbeel, and Sergey Levine. “Model-agnostic meta-learning for fast adaptation of deep networks.” ICML 2017

## 4 实验结果

MAML 要计算“梯度的梯度”，很慢。如果把后半部分换成直接算梯度，直接更新，而不是找“最小的梯度”，那么能快 33%，且性能掉的不多（但是这样和传统的机器学习梯度下降还是不同的？我有点乱了 【#TODO2】）。说明 MAML 的改进主要来自前半部分的参数更新（晕了）。

## 7 引用和拓展阅读

[2] What is Model-Agnostic Meta-learning (MAML) ? 是这篇文章的讲解，清晰易懂，需要翻墙

## 8 TODO

1. The MAML meta-gradient update involves a gradient through a gradient. Computationally, this requires an additional backward pass through f to compute Hessian-vector products, which is supported by standard deep learning libraries such as TensorFlow.
2. On MiniImagenet, we show a comparison to a first-order approximation of MAML, where these second derivatives are omitted. Note that the resulting method still computes the meta-gradient at the post-update parameter values θi′, which provides for effective meta-learning.
3. [3] 中我还不太能理解的句子：The meta-learning approach of both Reptile and MAML is to come up with an initialization for neural networks that is easily generalizable to similar tasks. This is different to “Learning to Learn by Gradient Descent by Gradient Descent” in which we weren’t learning an initialization but rather an optimizer. 为什么 meta-LSTM 的目标不是找一个好的 initial？