【论文笔记】Rethinking Few-Shot Image Classification: A Good Embedding is All You Need?

一、写在前面

本文是 MIT CSAIL & Google Research 在2020年关于 Few-Shot Learning的又一篇力作,受 ICLR 2020 的经典文章 A baseline for few-shot image classification 启发,提出了如下假设:

Embeddings are the most critical factor to the performance of few-shot learning/meta learning algorithms; better embeddings will lead to better few-shot testing performance (even with a simple linear classier)

二、主要内容

2.1 backgrounds

meta-learning 大致分为两个主要思路:

  • Optimization-based methods
  • Metric-based methods

few-shot classification tasks 也有两个思路:

  • meta-learning 方法
  • 学习好的表征 representation

2.2 related work

  • Metric-based meta-learning
    • Siamese neural networks for one-shot image recognition(ICML 2015)
    • Matching networks for one shot learning (NIPS 2016)
    • Prototypical networks for few-shot learning (NIPS 2017)
    • ...
  • Optimization-based meta-learning
    • Optimization as a model for few-shot learning (ICLR 2017)
    • MAML - Modelagnostic meta-learning for fast adaptation of deep networks. (ICML 2017)
    • ...
  • Meta-learning datasets
    • Omniglot
    • Meta-dataset
    • mini-ImageNet
    • ...

2.3 motivation

Few-shot classification中一直有这样一个问题:meta-learning 和 representation-learning 哪个表现更好。最近的文章发现仅用常规的模型做 fine-tuning,就可以达到与最好的meta-learning方法差不多的效果。因此本文探索在Few-shot classification中如何学习好的 representation。

三、文章方法

3.1 key idea

正如题目所说,A Good Embedding is All You Need。本文提出一个简单的few-shot classification baseline,学习 good learned representations 可以干掉一众 meta-learning algorithms

3.2 methods

具体怎么学习 good learned representations ?

  1. meta-training阶段,在组合训练集(可以理解为大规模数据集)上训练一个图像分类任务得到 embedding model,固定参数。
  1. meta-testing 阶段,将所有的 support images 和 query images 输入至上一步训练好的embedding model 得到对应的 embeddings (与组合训练集可以不同分布)
  2. 然后用support images embeddings 训练一个简单的线性模型 base learner
  3. 最后将 query images embeddings 输入至线性模型中得到分类结果


  • 值得注意的是,meta-testing 阶段中,每个task(个人理解是一组[support set, query set])都对 base learner 初始化再重新学习。
  • 此外,作者使用知识蒸馏(Knowledge distillation)得到了一个新的 embedding model ,获得了更好的效果 (2-3个点)

四、写在最后

这篇工作给 few-shot-classification打开了一个大门,验证了representation的重要性。此外,作者没有全盘否认meta-learning,只是说在 few-shot-classification任务中本文的方法比meta-learning好。

你可能感兴趣的:(【论文笔记】Rethinking Few-Shot Image Classification: A Good Embedding is All You Need?)