自Zero Shot Learning 提出以來，許多學者紛紛投入研究，而要真正實現ZSL需要解決兩個問題：第一個是獲取合適的語意描述向量;第二個問題則是建立一個合適的分類模型。然而目前大家致力於研究第二個問題，反而比較少關注第一個問題。上文所提到的語意描述向量大多是由原資料集所提供，該語意描述向量的產生主要由兩種方式產生：第一種是由該領域的專家定義;而第二種是透過NLP技術在維基百科中彙整(難度相當高)。而我提出一個方法，改善語意描述向量的產生方式，並且有以下幾個特點：第一是能夠量化特徵數值，第二是除了表現出"正"特徵，同時也帶有"負"特徵的表現。實驗時，我將模型所學習出來的語意描述向量置入其他人所設計的Zero Shot Learning Model，在Seen /Unseen /Generalized三種不同準確度下，數據都比使用原資料集所提供的語意描述向量好上許多，並且大幅提升。說明我所學習出來的語意描述向量，在分類上起了很大的幫助。

Since the introduction of Zero-Shot Learning, many scholars have invested in research, and to truly realize Zero-Shot Learning, two problems need to be solved: the first is to obtain a suitable semantic description vector or called class embedding; the second is to establish a suitable classification model. However, at present, everyone is devoted to studying the second question but pays less attention to the first question.
The class embedding mentioned above is mostly provided by the original data set. The generation of the class embedding is mainly generated in two ways: the first is defined by experts in the field; The second is to aggregate in Wikipedia through NLP technology. And I propose a method to improve the generation of the class embedding and has the following characteristics: 1. It can quantify the value of feature 2. In addition to showing the "positive" feature, it also has the performance of the "negative" feature. During the experiment, I replaced the class embedding learned from the P-Learning model with the class embedding provided in the data set with the Zero-Shot Learning Model designed by others. Under the three different accuracy levels of Seen / Unseen / Generalized, the accuracy is much better than that proposed in the original paper. It shows that the class embedding learned by P-Learning Model is very helpful in classification.

[1] C. H. Lampert, H. Nickisch, and S. Harmeling, "Learning to detect unseen object classes by between-class attribute transfer," 2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 951-958, doi: 10.1109/CVPR.2009.5206594.
[2] Kingma, D. P. & Welling, M. (2013). Auto-Encoding Variational Bayes (cite arxiv:1312.6114)
[3] A. Mishra, S. K. Reddy, A. Mittal and H. A. Murthy, "A Generative Model for Zero Shot Learning Using Conditional Variational Autoencoders," 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2018, pp. 2269-22698, doi: 10.1109/CVPRW.2018.00294.
[4] Y. Fu, T. M. Hospedales, T. Xiang and S. Gong, "Transductive Multi-View Zero-Shot Learning," in IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 37, no. 11, pp. 2332-2345, 1 Nov. 2015, doi: 10.1109/TPAMI.2015.2408354.
[5] W.-L. Chao, S. Changpinyo, B. Gong, and F. Sha, “An em- pirical study and analysis of generalized zero-shot learning for object recognition in the wild,” in ECCV, 2016.
[6] Y. Xian, B. Schiele, and Z. Akata, “Zero-shot learning-the good, the bad and the ugly,” in CVPR, 2017.
[7] M. Palatucci, D. Pomerleau, G. E. Hinton, and T. M. Mitchell, “Zero-shot learning with semantic output codes,” in NIPS, 2009.M. Palatucci, D. Pomerleau, G. E. Hinton, and T. M. Mitchell, “Zero-shot learning with semantic output codes,” in NIPS, 2009.
[8] R. Socher, M. Ganjoo, C. D. Manning, and A. Ng, “Zero- shot learning through cross-modal transfer,” in NIPS, 2013.R. Socher, M. Ganjoo, C. D. Manning, and A. Ng, “Zero- shot learning through cross-modal transfer,” in NIPS, 2013.
[9] A. Farhadi, I. Endres, D. Hoiem, and D. Forsyth, “Describing objects by their attributes,” in CVPR, 2009.
[10] V.KumarVerma,G.Arora,A.Mishra,andP.Rai,“General- ized zero-shot learning via synthesized examples,” in CVPR, 2018.
[11] M. Bucher, S. Herbin, and F. Jurie, “Generating visual rep- resentations for zero-shot classification,” in ICCV, 2017.
[12] A. Mishra, S. Krishna Reddy, A. Mittal, and H. A. Murthy, “A generative model for zero shot learning using conditional variational autoencoders,” in CVPR, 2018.
[13] Arash Vahdat, Jan Kautz. NVAE: A Deep Hierarchical Variational Autoencoder. IEEE Conference on Computer Vision and Pattern Recognition(CVPR), 2020.
[14] Y. Xian, C. H. Lampert, B. Schiele, and Z. Akata, “Zero- shot learning-a comprehensive evaluation of the good, the bad and the ugly,” TPAMI, 2018.
[15] Catherine Wah, Steve Branson, Peter Welinder, Pietro Per-ona, and Serge Belongie. The caltech-ucsd birds-200-2011dataset. 2011
[16] Genevieve Patterson and James Hays. Sun attribute database: Discovering, annotating, and recognizing scene attributes. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2751–2758. IEEE, 2012.
[17] Samir Bhattarai. New Plant Diseases Dataset. Kaggle.
[18] Latent Embedding Feedback and Discriminative Features for Zero-Shot Classification, Narayan, Sanath and Gupta, Akshita and Khan, Fahad Shahbaz and Snoek, Cees GM and Shao, Ling , ECCV, 2020
[19] Generalized zero-and few-shot learning via aligned variational autoencoders, Schonfeld, Edgar and Ebrahimi, Sayna and Sinha, Samarth and Darrell, Trevor and Akata, Zeynep, Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 8247—8255 , 2019
[20] Bernardino Romera-Paredes and Philip H. S. Torr. 2015. An embarrassingly simple approach to zero-shot learning. In Proceedings of the 32nd International Conference on International Conference on Machine Learning - Volume 37 (ICML'15). JMLR.org, 2152–2161.