Go to ICLR 2024 Conference homepageGeneralized Category Discovery with Hierarchical Label Smoothing
\textit{Generalized Category Discovery} seeks to cluster unidentified categories while simultaneously discerning known categories. Existing approaches predominantly rely on contrastive learning to produce distinctive embeddings for both labeled and unlabeled data. Yet, these techniques often suffer from dispersed clusters for unknown categories due to the lack of discriminative cues and a high rate of false negatives, thereby compromising the model’s ability to discriminate clusters effectively. To alleviate this problem, we introduce label smoothing as a hyperparameter that permits ‘forgivable mistakes’ when samples are closely related. We introduce a self-supervised cluster hierarchy, which allows us to control the strength of label smoothing to apply. By assigning pseudo labels to emerging cluster candidates and using these as ‘soft supervision’ for contrastive learning, we effectively combine the benefits of clustering-based learning and contrastive learning. The resulting method is applicable for both unsupervised and semi-supervised scenarios and we demonstrate state-of-the-art generalized category discovery performance on various fine-grained datasets.