why chose "Supervised Contrastive loss"

[hppy139]

Inspired by <Zhou et al., 2021>, in Section 3.4, you chose to do contrastive learning by using "Supervised Contrastive loss".
But in <Zhou et al., 2021>, I found that "Margin-based Contrastive Loss" may be the better choice.

So, I wonder for why?

[ddehun]

Hi, thank you for your interest in our work!

The decision to use SCL instead of Margin-based CL was primarily a design choice. SCL is simpler to implement due to having fewer hyperparameters, and it already provided reasonable performance gains. Based on this, we decided not to experiment further with other CL objectives. That said, I agree that exploring more advanced or alternative objectives could potentially enhance the overall performance of the models.

Thanks again!