Question about the regression problem

[Ma Junteng]

Hi, I have a question about the method of regression. In section 4.2 in your paper, why use kl-divergence loss between p(yu | xin) and the scaled score (y−vl)/(vu−vl) (loss = loss_fct(logits.view(-1, 2), labels)), but not cross entropy loss? The logits and labels here are both probability distribution on the 2 polarities.
Thanks in advance!

[Adam Fisch]

Thanks for the question. It should be possible to use either one.

[Ma Junteng]

Thanks for the question. It should be possible to use either one.

Thanks for the quick reply!