Questions Regarding Methodology Design

[Uoruichi]

First of all, we would like to thank the authors for their excellent work. However, we would like to raise the following questions:

1 How does aligning the denoised student model's features with the teacher's latent space features benefit the optimization of the student model?
2 During inference, the student model's features are not denoised. How does this approach improve the student model's performance?
3 How can this improvement be demonstrated through implementation?
4 How can the effectiveness of this approach be substantiated with data?