小樣本識別的提出，是為了利用少量的訓練資料，便能辨識全新的類別，並
省去資料蒐集與標記的成本。在本篇論文中，我們提出了一個可學習的正規度量
(Metric) 與一種整合度量學習中類原型學習與鄰近資料學習的方法，用以量測每
一對資料不相似的程度，來解決現今度量學習方法中的問題。而為了選出適合的
特徵以利於進行不相似程度量測，我們提出了一個特徵選取的作法，透過學習門
檻來過濾出有鑑別力的特徵。另一方面，考量到影像中的物體的尺度變化較大，
我們提出了一個多尺度特徵提取器以取出不同尺度的特徵，並利用訓練後的度量
(Metric)量測不同尺度的物體之間的關係。在 Omniglot 資料集與 miniImageNet 資料集中，實驗結果顯示整合現有方法對於現有的度量學習在效能上有所提升，實驗也顯示我們的方法在小樣本識別上，得到了與最先進方法近似的結果，尤其是在訓練樣本數量大於 1 的實驗，並且使用了相對較少的參數、減少了過擬合的可能。

Few-shot recognition aims to recognize novel classes with only a few training samples and alleviates the cost of data collection and labeling. In this thesis, we propose a deep metric and integrate existing metric learning approaches to compare the dissimilarity between each data pair. In addition, we propose a feature selection method by learning a threshold to select discriminative features. Considering that objects vary in scales, we propose a multi-scale feature extractor and include the extracted features in the learned metric to ensure the multi-scale property. Our experiments show that the integration of existing metric learning approaches improves performance on Omniglot dataset and the miniImageNet dataset. Furthermore, experimental results show that our model achieves competitive results to state-of-the-art methods, especially when the amount of training data is more than 1, while using much fewer parameters.

中文摘要 I
Abstract II
1. Introduction 1
2. Related Work 6
2.1 Feature Selection 6
2.2 Metric Learning 7
3. Proposed Method 10
3.1 Episodic Training 11
3.2 Multi-Scale Feature Extraction 12
3.3 Feature Selection 13
3.3.1. Threshold Learning 14
3.3.2. Relaxation of the Top-N feature Selection 15
3.4 Metric Learning 15
3.5 Training Loss 18
4. Experimental Results 20
4.1 Datasets 20
4.2 Implementation Details 21
4.3 Evaluation Scheme 22
4.4 Evaluation of Different Components 22
4.4.1 Evaluation of Metric Learning 24
4.4.2 Evaluation of Feature Selection 25
4.5 Comparison with Existing Methods 28
4.6 Discussion 31
4.6.1 Failure Cases 31
4.6.2 Limitation of Few-shot Learning 32
5. Conclusions 35
6. References 36

[1] Li and N. Vasconcelos. Multiple Instance Learning for Soft Bags via Top Instances. In CVPR, 2015.
[2] X. Cai, F. Nie and H. Huang. Exact Top-k Feature Selection via ℓ2,0-Norm Constraint. In IJCAI, 2013.
[3] T. Durand, T. Mordan, N. Thome and M. Cord. WILDCAT: Weakly Supervised Learning of Deep ConvNets for Image Classification, Pointwise Localization and Segmentation. In CVPR, 2017.
[4] G. Huang, D. Chen, T. Li, and F. Wu. Multi-scale dense networks for resource efficient image classification. In ICLR, 2018.
[5] O. Vinyals, C. Blundell, T. Lillicrap, K. Kavukcuoglu, and D. Wierstra. Matching networks for one shot learning. In NIPS, 2016.
[6] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, A. Karpathy Z. Huang, A. Khosla, M. Bernstein, A.C. Berg, and L. Fei-Fei. Imagenet large scale visual recognition challenge. IJCV, 115(3):211–252, 2015.
[7] B. M. Lake, R. Salakhutdinov, J. Gross, and J. B. Tenenbaum. One shot learning of simple visual concepts. In CogSci, 2011.
[8] C. Finn, P. Abbeel, and S. Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In ICML, 2017.
[9] S. Ravi and H. Larochelle. Optimization as a model for few-shot learning. In ICLR, 2017.
[10] B. Hariharan and R. Girshick. Low-shot Visual Recognition by Shrinking and Hallucinating Features. In ICCV, 2017.
[11] N. Mishra, M. Rohaninejad, X. Chen and P. Abbeel. A Simple Neural Attentitive Meta-learner. In ICLR, 2018
[12] G. Koch, R. Zemel, and R. Salakhutdinov. Siamese neural networks for one-shot image recognition. In ICML Workshop, 2015.
[13] J. Snell, K. Swersky, and R. S. Zemel. Prototypical networks for few-shot learning. In NIPS, 2017.
[14] B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba. Learning deep features for discriminative localization. In CVPR, 2016.
[15] V. Garcia and J. Bruna. Few-shot Learning With Graph Neural Networks. In ICLR, 2018.
[16] F. Sung, Y. Yang, L. Zhang, T. Xiang, P. H.S. Torr and T.M. Hospedales. Learning to Compare: Relation Network for Few-Shot Learning. In CVPR, 2018.
[17] T. Munkhdalai and H. Yu. Meta networks. arXiv preprint arXiv:1703.00837, 2017.
[18] A. Santoro, S. Bartunov, M. Botvinick, D. Wierstra, and T. Lillicrap. Meta-learning with memory-augmented neural networks. In ICML, 2016.
[19] L. Fei-Fei, R. Fergus, and P. Perona. One-shot learning of object categories. TPAMI, 2006.
[20] T. Durand, N. Thome, and M. Cord. MANTRA: Minimum Maximum Latent Structural SVM for Image Classification and Ranking. In ICCV, 2015.
[21] T. Durand, N. Thome, and M. Cord. WELDON: Weakly supervised learning of deep convolutional neural networks. In CVPR, 2016.
[22] M. Oquab, L. Bottou, I. Laptev, and J. Sivic. Is object localization for free? Weakly-supervised learning with convolutional neural networks. In CVPR, 2015.
[23] C. Sun, M. Paluri, R. Collobert, R. Nevatia, and L. Bourdev. ProNet: Learning to Propose Object-Specific Boxes for Cascaded Neural Networks. In CVPR, 2016.
[24] F. X. Yu, D. Liu, S. Kumar, T. Jebara, and S.-F. Chang. ∝svm for learning with label proportions. In ICML, 2013.
[25] S. N. Parizi, A. Vedaldi, A. Zisserman, and P. F. Felzenszwalb. Automatic discovery and optimization of parts for image classification. In ICLR, 2015.
[26] R. Arandjelovic, P. Gronat, A. Torii, T. Pajdla, and J. Sivic. NetVLAD: CNN architecture for weakly supervised place recognition. In CVPR, 2016.
[27] Pytorch. URL https://github.com/pytorch/pytorch.
[28] H. Edwards and A, Storkey. Towards a neural statistician. arXiv preprint arXiv:1606.02185, 2016.
[29] L. Kaiser, O. Nachum, A. Roy and S. Bengio. Learning to remember rare events. arXiv preprint arXiv:1703.03129, 2017.
[30] Z. Li, F. Zhou, F. Chen and H. Li. Meta-sgd: Learning to learn quickly for few shot learning. arXiv preprint arXiv:1707.09835, 2017.