What the difference between GCD tasks and Open-world semi-supervised learning tasks？

[jinweiisgreat]

Thanks to this repository for its contribution to the NCD field.

But I’m still not very clear about the difference between GCD tasks and Open-world semi-supervised learning tasks. Maybe they are the same task.

Thank you！

[ColinTr]

There is no difference, both terms refer to the same problem where you try to (1) classify the known classes and (2) cluster the novel classes. The field is recent and there are conflicting terms for the same problems. Another paper even called it "Open Set Domain Adaptation".

[FanZhichen]

These two problem settings are introduced from different perspectives, but they are the same thing. In NCD, labeled and unlabeled data are class-disjoint, so what if some unlabeled data come from known classes (i.e., GCD). Meanwhile, in SSL, labeled and unlabeled data share the same output space, and what if some unlabeled data come from novel classes (i.e., open-world setting).

[jinweiisgreat]

I see! Thank you all!