Possible Confusion Regarding Figure 2

[Inpyo-Hong]

Thank you for your outstanding work on adaptive data-free quantization.
I found your paper both insightful and innovative. However, while reviewing figure 2 in your paper, I noticed a potential inconsistency regarding the interpretation of $H'_{\text{info}}$ and its relationship to overfitting and underfitting.

Based on your paper's definitions and discussions:

• $H'_{\text{info}}$ represents the uncertainty of a generated sample.
• Higher $H'_{\text{info}}$ (closer to 1) corresponds to higher uncertainty and underfitting.
• Lower $H'_{\text{info}}$(closer to 0) corresponds to lower uncertainty and overfitting.

This relationship is clearly supported by your explanation in the text:

"Encouraging the sample with lowest adaptability (i.e., largest $H'_{\text{info}}$) may lead to ..."

This directly implies that larger $H'_{\text{info}}$ values are associated with lower adaptability, which aligns with underfitting.

   

However, Figure 2 appears to indicate the following:

• $H'_{\text{info}}$ = 1 (high uncertainty) corresponds to overfitting.
• $H'_{\text{info}}$ = 0 (low uncertainty) corresponds to underfitting.

This interpretation seems to contradict the definitions and explanations provided in your paper. According to these definitions, it appears that:

• $H_{\text{nor}}$ = 1 should correspond to overfitting (low uncertainty, high adaptability).
• $H_{\text{nor}}$ = 0 should correspond to underfitting (high uncertainty, low adaptability).

   

Could you kindly confirm if my understanding is correct? Is Figure 2 incorrectly labeling the overfitting and underfitting regions, or is there an alternative interpretation I may have missed?

Thank you for your time and for addressing this question. I appreciate your valuable contributions to this field.