[Few-shot Classification]Review:(ICLR 2019）A Close Look at Few-shot Classification

前言
本次的Review，我将按照[How to read a paper]的顺序进行阅读。
[open review]
[code]

1. General Information

文章是对Few-shot classification的一个综述性质的文章，但是在其基础上还提出了自己的模型，并对以前的方法进行了统一的标准。

文章的主要贡献如下：

• 对现有的few-shot classification算法做了公平性的比较分析，结果显示更深的backbone明显的降低了不同模型方法的表现效果区别，也就是说，当backbone很深的时候，不同的模型表现效果差不多；
• 在CUB数据集和mini-ImageNet数据集上，经过作者修改过得baseline和其他STOA的模型相比，实现了更好的表现。
• 此外，作者还针对few-shot Classification算法设定了一个新的实验场景，用来衡量跨域泛化能力。

2. Related

Baseline

作者提出的baseline模型来自下面两篇文章

• Victor Garcia and Joan Bruna. Few-shot learning with graph neural networks. In Proceedings of the
  International Conference on Learning Representations (ICLR), 2018

• Hang Qi, Matthew Brown, and David G Lowe. Low-shot learning with imprinted weights. In
  Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018.

Initialization based methods

Learning to fine-tune
学习一个好的初始化模型，对于新的数据集做"fine-tune"

• Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the International Conference on Machine Learning (ICML), 2017 [paper] [TensorFlow]

• Alex Nichol and John Schulman. Reptile: a scalable metalearning algorithm. arXiv preprint arXiv:1803.02999, 2018

• Andrei A Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. Meta learning with latent embedding optimization. In Proceedings of the International Conference on Learning Representations (ICLR), 2019

Learning to optimizer

• Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In Proceedings of the International Conference on Learning Representations (ICLR), 2017

• Tsendsuren Munkhdalai and Hong Yu. Meta networks. In Proceedings of the International Conference on Machine Learning (ICML), 2017

Distance metric learning based methods

Learning to cmpare

• Gregory Koch, Richard Zemel, and Ruslan Salakhutdinov. Siamese neural networks for one-shot image recognition. In Proceedings of the International Conference on Machine Learning Workshops (ICML Workshops), 2015.

  • 孪生网络计算相似度

• Oriol Vinyals, Charles Blundell, Tim Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. In Advances in Neural Information Processing Systems (NIPS), 2016

  • cosine 相似度求距离

• Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In
  Advances in Neural Information Processing Systems (NIPS), 2017.

  • Euclidean distance 求每个类别的中心表示距离

• Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018

  • Relation Network, 训练一个好的分类器

• Luca Bertinetto, Joao F Henriques, Philip HS Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. In Proceedings of the International Conference on Learning Representations (ICLR), 2019

• Victor Garcia and Joan Bruna. Few-shot learning with graph neural networks. In Proceedings of the International Conference on Learning Representations (ICLR), 2018

Hallucination based methods

Learning to augment

• Bharath Hariharan and Ross Girshick. Low-shot visual recognition by shrinking and hallucinating features. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017.

• Antreas Antoniou, Amos Storkey, and Harrison Edwards. Data augmentation generative adversarial networks. In Proceedings of the International Conference on Learning Representations Workshops (ICLR Workshops), 2018.

• Yu-Xiong Wang, Ross Girshick, Martial Hebert, and Bharath Hariharan. Low-shot learning from
  imaginary data. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018

Domain adaptation

• Nanqing Dong and Eric P Xing. Domain adaption in one-shot learning. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, 2018.

解释reduce intra-class

解释为何欧几里得距离为何可以reduce intra-class：

• Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In Advances in Neural Information Processing Systems (NIPS), 2017

• Oriol Vinyals, Charles Blundell, Tim Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. In Advances in Neural Information Processing Systems (NIPS), 2016

3. 应用普遍的Few-shot Classification算法

作者首先根据前人的工作设计了两个baseline网络，分别是baseline和baseline++，两个网络的训练流程是一样的，在训练阶段采用基础类别数据训练一个特征提取器和分类器，在fine-tune阶段，固定特征提取器的参数，采用新的类别样本对分类器进行fine-tune；两者的区别在于Baseline的分类器采用的是Linear laye；而Baseline++采用的分类器是cosine distance和对每个类的weight vector用来减少intra-class variations。

对此复现了四种常见的Few-shot Classification算法， 分别是:

• MatchingNet
• ProtoNet
• RelationNet
• MAML

作者将以上四种方法归类为Meta-learning few-shot classification algorithms, 这类算法的流程是，首先在meta-train阶段，support set $S_b$和query set $Q_b$采用Episode的训练方法训练一个meta-learning classifier $M(.|Sb)$ ;
在meta-testing阶段，Novel support set $S_n$ 训练分类器$M(.|S_n)$ 用来预测新类别中的目标。对不不同的meta-learning算法，他们的区别主要是在分类器$M(.|S)$的设计上有区别。

4. 实验结果

文中作者设计了3个场景：

• generic object recognition(mini-ImageNet)
• fine-grained image classification(CUB-200-2011)
• cross-domain adaptation(mini-ImageNet->CUB-200-2011)

结果表明，随着backbone的加深，不同算法之间的差距在逐渐缩小，说明更深的网络可以reduce intra-class variation。