少樣本開放資料識別 (Few­shot open­set recognitio, FSOR) 是一項具有挑戰性的任務，旨在僅憑有限數量的標註資料辨識已知類別的樣本（正樣本），同時檢測不屬於任何已知類別的樣本（負樣本）。這個問題之所以非常困難是因為模型必須從少量標註樣本中學習泛化，並將其與無限數量的潛在負樣本區分開來。現有方法試圖建立不同類別的原型 (prototype)，並使用閾值 (threshold) 建立分類器以檢測負樣本。然而，此方法存在閾值設定問題，且結果常常不穩定且難以滿足要求。
本研究提出了一種創新的方法，稱為整體正原型 (overall positive prototype)，可顯著提升 FSOR 的性能。與其關注分散在特徵空間中且難以準確描述的負樣本不同，我們提議建立一個整體正原型，作為相對較小鄰域中正樣本的一個整體表示。通過計算樣本與整體正原型之間的距離，我們能夠有效地將其分類為正樣本或負樣本。
我們的方法簡單而創新，在準確性 (accuracy) 和接 ROC 曲線下面積 (AUROC) 方面提供了很好的 FSOR 性能，超越現有方法。這凸顯了整體正原型在改善少樣本開放資料識別性能方面的有效性。

Few­shot open­set recognition (FSOR) is a challenging task that involves recognizing samples belonging to known classes with only a limited number of annotated instances, while also detecting samples that do not belong to any known class. Existing approaches address this problem by constructing prototypes for different classes and employing a threshold­based classifier to identify negative samples. However, these methods suffer from the issue of threshold setting and often fail to consistently achieve satisfactory results.
In this thesis, we present a novel approach called the overall positive prototype, which significantly enhances the performance of FSOR. Instead of focusing on negative samples scattered throughout the feature space and hardly to be described, we propose constructing an overall positive prototype that serves as a coherent representation for positive samples located in a relatively smaller neighborhood. By evaluating the distance between a query sample and the overall positive prototype, we can effectively classify it as positive or negative.
Our approach demonstrates its simplicity and innovation by achieving state­of­the­art performance in FSOR, surpassing existing methods in terms of accuracy and AUROC. This highlights the effectiveness of the overall positive prototype in improving the performance of few­shot open­set recognition.

[1] Abhijit Bendale and Terrance E. Boult. Towards open set deep networks. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 1563–1572, 2016.
[2] Luca Bertinetto, Joao F. Henriques, Philip H.S. Torr, and Andrea Vedaldi. Meta­learning with differentiable closed­form solvers. In Proceedings of International Conference on Learning Representations, 2019.
[3] Malik Boudiaf, Etienne Bennequin, Myriam Tami, Antoine Toubhans, Pablo Piantanida, Celine Hudelot, and Ismail Ben Ayed. Open­set likelihood maximization for few­shot learning. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023.
[4] Malik Boudiaf, Ziko Imtiaz Masud, Jerome Rony, Jose Dolz, Pablo Piantanida, and Ismail Ben Ayed. Transductive information maximization for few­shot learning. In Proceedings of Neural Information Processing Systems, 2020.
[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jegou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self­supervised vision transformers. In Proceedings of International Conference on Computer Vision, pages 9650–9660, 2021.
[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of International Conference on Learning Representations, 2021.
[7] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model­agnostic meta­learning for fast adaptation of deep networks. In Proceedings of International Conference on Machine Learning, pages 1126–1135. PMLR, 2017.
[8] ZongYuan Ge, Sergey Demyanov, Zetao Chen, and Rahil Garnavi. Generative openmax for multi­class open set classification. In Proceedings of British Machine Vision Conference, 2017.
[9] Spyros Gidaris and Nikos Komodakis. Dynamic few­shot visual learning without forgetting. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 4367–4375, 2018.
[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016.
[11] Shell Xu Hu, Da Li, Jan Stuhmer, Minyoung Kim, and Timothy M. Hospedales. Pushing the limits of simple pipelines for few­shot learning: External data and fine­tuning make a difference. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 9068–9077, 2022.
[12] Yuqing Hu, Vincent Gripon, and Stephane Pateux. Leveraging the feature distribution in transfer­based few­shot learning. In Proceedings of International Conference on Artificial Neural Networks, 2021.
[13] Shiyuan Huang, Jiawei Ma, Guangxing Han, and Shih­Fu Chang. Task­adaptive negative envision for few­shot open­set recognition. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7171–7180, 2022.
[14] Minki Jeong, Seokeon Choi, and Changick Kim. Few­shot open­set recognition by transformation consistency. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12566–12575, 2021.
[15] Zhenguo Li, Fengwei Zhou, Fei Chen, and Hang Li. Meta­sgd: Learning to learn quickly for few­shot learning. arXiv preprint arXiv:1707.09835, 2017.
[16] Bo Liu, Hao Kang, Haoxiang Li, Gang Hua, and Nuno Vasconcelos. Few­shot open­set recognition using meta­learning. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020.
[17] Jinlu Liu, Liang Song, and Yongqiang Qin. Prototype rectification for few­shot learning. In Proceedings of European Conference on Computer Vision, 2020.
[18] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X. Yu. Large­scale long­tailed recognition in an open world. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 2537–2546, 2019.
[19] Lawrence Neal, Matthew Olson, Xiaoli Fern, Weng­Keen Wong, and Fuxin Li. Open set learning with counterfactual images. In Proceedings of European Conference on Computer Vision, pages 613–628, 2018.
[20] Alex Nichol, Joshua Achiam, and John Schulman. On first­order meta­learning algorithms. arXiv preprint arXiv:1803.02999, 2018.
[21] Boris N. Oreshkin, Pau Rodriguez, and Alexandre Lacoste. Tadam: Task dependent adaptive metric for improved few­shot learning. In Proceedings of Advances in Neural Information Processing Systems, 2019.
[22] Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B. Tenenbaum, Hugo Larochelle, and Richard S. Zemel. Meta­learning for semisupervised few­shot classification. In Proceedings of International Conference on Learning Representations, 2018.
[23] Walter J Scheirer, Anderson de Rezende Rocha, Archana Sapkota, and Terrance E Boult. Toward open set recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(7):1757–1772, 2012.
[24] Jake Snell, Kevin Swersky, and Richard S. Zemel. Prototypical networks for few­shot learning. In Proceedings of Neural Information Processing Systems, 2017.
[25] Nan Song, Chi Zhang, and Guosheng Lin. Few­shot open­set recognition using background as unknowns. In Proceedings of ACM International Conference on Multimedia, pages 5970–5979, 2022.
[26] Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip H.S. Torr, and Timothy M. Hospedales. Learning to compare: Relation network for few­shot learning. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 1199–1208, 2018.
[27] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open­set recognition: A good closed­set classifier is all you need. arXiv preprint arXiv:2110.06207, 2021.
[28] Olivier Veilleux, Malik Boudiaf, Pablo Piantanida, and Ismail Ben Ayed. Realistic evaluation of transductive few­shot learning. In Proceedings of Neural Information Processing Systems, 2021.
[29] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Proceedings of Advances in Neural Information Processing Systems, 2016.
[30] Yan Wang, Wei­Lun Chao, Kilian Q. Weinberger, and Laurens van der Maaten. Simpleshot: Revisiting nearest­neighbor classification for few­shot learning. In arXiv:1911.04623, 2019.
[31] Yikai Wang, Chengming Xu, Chen Liu, Li Zhang, and Yanwei Fu. Instance credibility inference for few­shot learning. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, 2020.
[32] Han­Jia Ye, Hexiang Hu, De­Chuan Zhan, and Fei Sha. Few­shot learning via embedding adaptation with set­to­set functions. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 8808–8817, 2020.
[33] Baosheng Yu and Dacheng Tao. Deep metric learning with tuplet margin loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6490–6499, 2019.
[34] Da­Wei Zhou, Han­Jia Ye, and De­Chuan Zhan. Learning placeholders for open­set recognition. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 4401–4410, 2021.
[35] Imtiaz Masud Ziko, Jose Dolz, Eric Granger, and Ismail Ben Ayed. Laplacian regularized few­shot learning. In Proceedings of International Conference on Machine Learning, 2020.