Overview of Zero-Shot Learning - 1.3

Controbutions and Article Organization .

Controbutions:

• Although there are surveys on transfer learning , and particularly on heterogeneous transfer learning . they do not cover the topic of zero-shot lering with sufficient depth.
• To the best of our knowledge, only a few attempts have been made for literature review on zero-shot learning .
  -In [136] Based on how the Feature space and Semantic space are related, this article categorized the zero-shot learning methods into three categories:
  • one- order transformation approaches
  • two- order transformation approaches
  • high -order transformation appraches .

As the number of related works reviewed by this article is limited , this categorization is also limited.
Many existing methods do not belong to any of these categories.
On The other hand , in this article , just a brief introduction of some semantic spaces is given .
No formal categorization of existing semantic spaces is provided.

In [155], A categorization of 16 methods in zero-shot learning is given .
The criteria of method categorization are not given .
No summarization of these semantic spaces is given .

• IN [43], They use a unified "Embedding model and recognition in embedding space " to summarize the existing methods in zero-shot learning .

Summarize our contribetion

• As shown in Fig.2 , we provide a hierarchical categorization of existing methods in zero-shot learning .
  • First categorize methods ino two general categories based on the aim to get the classifiers for the unseen classes directly or aim to get the instances of the unseen classes.
  • Then , methods in each general category are further cagegorized . we provide a more comprehensive perspective for readers to understand the existing zero-shot learning , and select suitable ones for the application scenarios they encounter .
  • we provide a formal classification and definition of different learning settings in zero-shot