search

woshixuhao / semantic-explainable-AI

The codes and dataset for the semantic explainable AI (S-XAI)
14 stars 4 forks source link

questions about researches for S-XAI #2

Open xwhkkk opened 1 year ago

xwhkkk commented 1 year ago

Hello! Thanks for sharing your great job. I want to ask some questions.

  1. Can 'the common traits' mentioned in the paper be understood as the most discriminative parts?; Is it similar to the idea of the Class activation map (CAM)? e.g., By looking for the most discriminative feature of a cat to determine its category.
  2. For the application of trustworthiness assessment, could the S-XAI method be used in object detection or instance segmentation tasks?
woshixuhao commented 1 year ago

Thanks for your interest or our work! Here is my opinon.

  1. Yes, the 'common traits' are essentially the most discriminative parts of each class. However, the main idea is different from CAM. From my opinon, the CAM (e.g., Grad-CAM method) extacts the most discriminative parts in each single picture. For example, a picture with both a cat and a dog. CAM helps to visualize the corresponding parts that activate certain class mostly. In contrast, the 'common traits' are extracted from a large number of samples of the same class. In other words, they refer to the most discriminative parts of certain class (e.g., cats, dogs).
  2. It is a good idea, however, there may be some challenges that should be handled. As discussed in the Discussion, the dominant challenge is that the extraction of semantic space needs manually spliting the semantic part. Meanwhile, although we can obtain the semantic probability (i.e., distingush whether there are semantic concepts), the location can not be determinted. Further investigation may need to solve these problem.            Best regards.             Hao Xu          Peking university

------------------ 原始邮件 ------------------  发件人: "woshixuhao/semantic-explainable-AI" @.>; 发送时间: 2022年12月21日(星期三) 中午1:39 @.>; @.***>; 主题: [woshixuhao/semantic-explainable-AI] questions about researches for S-XAI (Issue #2)

Hello! Thanks for sharing your great job. I want to ask some questions.

Can 'the common traits' mentioned in the paper be understood as the most discriminative parts?; Is it similar to the idea of the Class activation map (CAM)? e.g., By looking for the most discriminative feature of a cat to determine its category.

For the application of trustworthiness assessment, could the S-XAI method be used in object detection or instance segmentation tasks?

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.Message ID: @.***>