ICLR 2019:A CLOSER LOOK AT FEW-SHOT CLASSIFICATION （ACK-FSC）

下方↓公众号后台回复“ACKFSC”，即可获得论文原文电子资源。

内容提要

本文工作

we present：

• a consistent comparative analysis of several representative few-shot classification algorithms, with results showing that deeper backbones significantly reduce the performance differences among methods on datasets with limited do- main differences(对几种具有代表性的few-shot分类算法进行了一致的比较分析，结果表明，在具有有限的域差异的数据集上，更深层次的骨干网络显著减少了方法之间的性能差异)
• a modified baseline method that surprisingly achieves com- petitive performance when compared with the state-of-the-art on both the mini- ImageNet and the CUB datasets（一种改进的基线方法，令人惊讶地在与最先进的mini- ImageNet和CUB数据集上获得具有竞争性的性能）
• a new experimental setting for evaluating the cross-domain generalization ability for few-shot classification algorithms. （提出了一个用于评估few-shot分类算法跨域泛化能力的新的实验设置。）

现有工作

promising direction of few-shot clas- sification

• meta-learning paradigm where transfer- able knowledge is extracted and propagated from a collection of tasks to prevent overfitting and improve generalization. （元学习范式，从任务集合中提取和传播可迁移的知识，以防止过拟合和改进泛化）
  • model initialization based methods //Ravi & Larochelle (2017); Finn et al. (2017),
  • metric learning methods // Vinyals et al. (2016); Snell et al. (2017); Sung et al. (2018),
  • hallucination based methods // Antoniou et al. (2018); Hariharan & Girshick (2017);
• directly predicting the weights of the classifiers for novel classes. (直接预测新类别分类器的权重。)

现有方法局限性

• there are two main challenges that prevent us from making a fair compari- son and measuring the actual progress.（有两个主要的挑战妨碍我们对实际进展进行公平的比较和衡量）
  • the discrepancy of the implementation details among multiple few-shot learning algorithms obscures the relative performance gain（不同的few-shot学习算法在实现细节上的差异掩盖了相对性能增益）
  • while the current evaluation focuses on recognizing novel class with limited training examples, these novel classes are sampled from the same dataset. The lack of domain shift between the base and novel classes makes the evaluation scenarios unrealistic （虽然目前的评估侧重于在有限的训练实例下识别新类，但这些新类都是从相同的数据集中采样。基础类和新类之间缺乏域转换，这使得评估场景不现实）

文章贡献

• We provide a unified testbed for several different few-shot classification algorithms for a fair comparison. Our empirical evaluation results reveal that the use of a shallow backbone commonly used in existing work leads to favorable results for methods that explicitly reduce intra-class variation. Increasing the model capacity of the feature backbone reduces the performance gap between different methods when domain differences are limited.（我们为几种不同的few-shot分类算法提供了一个统一的测试平台，以进行公平的比较。我们的实证评价结果显示，在现有工作中常用的浅层主干的使用，为明确减少类内变异的方法带来了良好的结果。在域差异有限的情况下，通过增加feature backbone的模型容量可以减小不同方法之间的性能差距。）
• We show that a baseline method with a distance-based classifier surprisingly achieves competitive performance with the state-of-the-art meta-learning methods on both mini-ImageNet and CUB datasets. （我们的研究表明，在微型imagenet和CUB数据集上，具有基于距离分类器的基线方法出人意料地在最先进的元学习方法中获得了具有竞争力的性能。）
• We investigate a practical evaluation setting where base and novel classes are sampled from dif- ferent domains. We show that current few-shot classification algorithms fail to address such do- main shifts and are inferior even to the baseline method, highlighting the importance of learning to adapt to domain differences in few-shot learning.（我们研究了一个实际的评估设置，基类和新类的样本来自不同的领域。我们发现，目前的few-shot分类算法未能解决这种do- main变换，甚至不如基线方法，这凸显了在few-shot学习中，学习适应领域差异的重要性。）

方法详解

实验结果

扩展解析

更多扩展解析 和 论文原文 见： xxxx