The LsrGAN does not use the image feature of unseen classes while training

[dyabel]

As the title

[WilliamYi96]

Yes, actually this should be semantic features instead of image features. I've fixed the typos. But it will not change the fact that this is a transductive zero-shot learning setting.

[dyabel]

I believe transductive zero-shot means using the image data of unseen classes. And semantic features of unseen classes should be known in standard ZSL and GZSL setting according to the literatures I have read.

[WilliamYi96]

Actually, for zero-shot learning, there can be four different settings:

1. Inductive zero-shot learning (Seen set)
2. Semantic transductive zero-shot learning (Seen set + labeled unseen attributes)
3. Feature transductive zero-shot learning (Seen set + unlabeled visual features)
4. General transductive zero-shot learning (Seen set + labeled unseen attributes + unlabeled visual features)

I think the conflict mainly comes from item I and item II. We claim pure zero-shot learning should not access unseen data. Otherwise, if we use the labeled semantic features, the searching space of unseen classes will be highly constrained and the generalization ability of a model will be highly decreased. In general, it's true we can have better performance and understand seen/unseen relationships better with accessing unseen data and most recent papers are following this setting.

I suggest you having a look at the following inductive zero-shot learning papers if you want:

1. GAZSL: Yizhe Zhu, Mohamed Elhoseiny, Bingchen Liu, Xi Peng, Ahmed Elgammal. "A Generative Adversarial Approach for Zero-Shot Learning From Noisy Texts".
2. CIZSL: Mohamed Elhoseiny, Mohamed Elfeki. Creativity Inspired Zero-Shot Learning.
3. CN-ZSL: Class Normalization for Zero Shot Learning. Ivan Skorokhodov, Mohamed Elhoseiny.
4. GRaWD: Imaginative Walks: Generative Random Walk Deviation Loss for Improved Unseen Learning Representation. Mohamed Elhoseiny, Divyansh Jha, Kai Yi, Ivan Skorokhodov.
5. Mancini, M., Akata, Z., Ricci, E., & Caputo, B. (2020). Towards recognizing unseen categories in unseen domains.

Let me know if you have any further concerns.

[dyabel]

Actually, for zero-shot learning, there can be four different settings:

1. Inductive zero-shot learning (Seen set)
2. Semantic transductive zero-shot learning (Seen set + labeled unseen attributes)
3. Feature transductive zero-shot learning (Seen set + unlabeled visual features)
4. General transductive zero-shot learning (Seen set + labeled unseen attributes + unlabeled visual features)

I think the conflict mainly comes from item I and item II. We claim pure zero-shot learning should not access unseen data. Otherwise, if we use the labeled semantic features, the searching space of unseen classes will be highly constrained and the generalization ability of a model will be highly decreased. In general, it's true we can have better performance and understand seen/unseen relationships better with accessing unseen data and most recent papers are following this setting.

I suggest you having a look at the following inductive zero-shot learning papers if you want:

1. GAZSL: Yizhe Zhu, Mohamed Elhoseiny, Bingchen Liu, Xi Peng, Ahmed Elgammal. "A Generative Adversarial Approach for Zero-Shot Learning From Noisy Texts".
2. CIZSL: Mohamed Elhoseiny, Mohamed Elfeki. Creativity Inspired Zero-Shot Learning.
3. CN-ZSL: Class Normalization for Zero Shot Learning. Ivan Skorokhodov, Mohamed Elhoseiny.
4. GRaWD: Imaginative Walks: Generative Random Walk Deviation Loss for Improved Unseen Learning Representation. Mohamed Elhoseiny, Divyansh Jha, Kai Yi, Ivan Skorokhodov.
5. Mancini, M., Akata, Z., Ricci, E., & Caputo, B. (2020). Towards recognizing unseen categories in unseen domains.

Let me know if you have any further concerns.

Thank you for your explanation， I will take a look at these papers.