BiGCD: A Bias-aware Framework for Generalized Category Discovery

Abstract

Generalized Category Discovery (GCD) requires a model to recognize labeled seen classes while discovering unlabeled novel classes from partially annotated data. This paper argues that a key difficulty comes from bias that accumulates during optimization. Two effects appear repeatedly in practice: feature response imbalance, where representations concentrate on a few high-activation dimensions, and prediction confidence shift, where the classifier becomes too confident on seen classes and pulls novel samples into known categories. To counter these effects, this paper proposes BiGCD, a bias-aware framework that regularizes both representations and predictions. BiGCD applies a Structural Suppression Module to dampen dominant feature dimensions and a Label Suppression Module to prevent overly peaked output distributions, and it activates the label-side regularization after a warm-up period to keep training stable. The modules are lightweight and can be plugged into existing GCD pipelines without adding learnable parameters. Experiments on CIFAR10, CIFAR100, Stanford Cars, FGVC-Aircraft, and Herbarium19 show consistent gains over strong baselines, with especially large improvements on seen-class recognition in fine-grained settings. Ablation studies further show that the two modules contribute independently and work better together.