Table 1.

Comparison of different problem settings

Problem setting	Label data	Unlabeled data
		Known classes	Unknown classes
Supervised learning	✓	✗	✗
Semi-supervised learning	✓	✓	✗
Robust Semi-supervised learning	✓	✓	Reject
Novel Class Discovery	✓	✗	✓
Generalized Category Discovery	✓	✓	✓

In the supervised learning setting, models are trained exclusively on labeled samples and they are only able to classify data (both labeled and unlabeled) into known classes. For semi-supervised and robust semi-supervised learning, the model leverages both labeled and unlabeled data, with the latter only classifying the known and rejecting the novel classes. Novel Class Discovery assumes that only novel classes exist in the unlabeled set and it is unable to re-discover known classes. Generalized Category Discovery aims to generalize Novel Class Discovery to further recognize the known classes in the unlabeled set.