Non-Parametric Few Shot Learning

Non-parametric Methods

Key idea: Use non-parametric learner.

Siamese networks

利用孪生网络预测两个图片是否是同一个类别:

Matching Networks

原本元训练是二分类问题,元测试是多分类问题,可以通过在高维学习到空间进行最近邻判断的方法将两者结合起来。 From Matching networks for one-shot learning.

Prototypical networks

From Prototypical networks for few-shot learning.

More Complex Relations

Properties of Meta-Learning Algorithms

Computation Graph Perspective

Algorithmic Properties Perspective

三种方法的具体区别:

Note: Cover Picture