Debiased Prototypical Learning Improves Generalized Category Discovery

Generalized Category Discovery (GCD) aims to discover novel and seen categories using knowledge from labeled samples. Existing methods like SimGCD showed a bias towards predicting seen categories. We identify that the primary source of misclassification in these methods is the incorrect categorization of unseen categories as seen categories. To mitigate this bias, we introduce a novel approach that leverages labels generated via k-means++ for each training example to compute dynamic prototypes for each mini-batch, providing more balanced supervision. Additionally, we incorporate prototypical contrastive learning loss that performs contrastive learning on these prototypes, reducing the bias towards seen categories and facilitating the learning of unbiased representations. Extensive experimental results on various GCD benchmarks reveal that our approach not only reduces the seen category bias in representation learning but also sets a new state-of-the-art performance across all tested benchmarks.