Unknown Sample Selection and Discriminative Classifier Learning for Generalized Category Discovery

Traditional supervised techniques rely on labeled data, which are not available for unknown classes. Generalized Category Discovery (GCD) aims to categorize data into both known and unknown classes. We proposed a method, Unknown Sample Selection and Discriminative Classifier Learning for GCD (USSDCL). Firstly, we map the features to a linear subspace to identify dimensions that best represent the semantic differences among the known classes and introduce a neighborhood relevance score to assess the consistency of labels among neighbors. Using this score, we can efficiently distinguish between known, unknown, and uncertain samples. Next, we leverage both the known samples, pseudo-labeled unknown samples, and the uncertain samples to train a discriminative classifier. This classifier incorporates both a cross-entropy loss and an uncertainty augmentation loss to encourage closer predictions between uncertain samples and their augmentations, enhancing the classifier’s discriminative capability. Our model exhibits enhanced performance in classifying both known and unknown samples.