yuhui-zh15 / drml Goto Github PK

Dear Yuhui,

Thank you for your great work. It is very interesting.

In section 2.1, you mention that the modality gap could be closed by subtracting the mean, x-E[X] and y-E[Y]. Could you share the mean image emb E[X] and text emb E[Y] with me? Which version of the CLIP model you are using?

Thank you for your help.

Best Wishes,

Zongze

Hi,

Thanks for your excellent work.
I have some questions. Would you like to give me some indications?

1. E_(x)[x-x^(T)g^(')g^(')], why you use this term as the center of x, is there exist a geometric meaning?
2. In Proposition A.1, you suppose that gap g is orthogonal to the span of image features, and image features have zero means in the subspace orthogonal to g, why do you explain that it is the projected x in the subspace orthogonal to g? I am confused about that. why there is a projection rather than the original x as you supposed?