Unleashing the Potential of Model Bias for Generalized Category Discovery

TL;DR

This work tackles Generalized Category Discovery by reframing model bias toward known categories as a learning signal for identifying novel categories. It introduces Self-Debiasing Calibration (SDC), a classifier-based framework that calibrates logits via two logit-adjustment modules—Category Bias Mitigation and Category Confusion Mitigation—guided by an entropy-based weighting, and uses Sinkhorn-Knopp to generate high-quality pseudo-labels for unlabeled data. The approach leverages a biased pre-trained model to derive bias representations and transfer knowledge between related categories, enabling online inference and improved identification of novel categories. Experiments on BANKING, HWU64, and CLINC demonstrate consistent gains over SOTA methods, with notable improvements in novel-category accuracy and balanced performance across known and novel classes, confirming the practical utility of bias-informed GCD.

Abstract

Generalized Category Discovery is a significant and complex task that aims to identify both known and undefined novel categories from a set of unlabeled data, leveraging another labeled dataset containing only known categories. The primary challenges stem from model bias induced by pre-training on only known categories and the lack of precise supervision for novel ones, leading to category bias towards known categories and category confusion among different novel categories, which hinders models' ability to identify novel categories effectively. To address these challenges, we propose a novel framework named Self-Debiasing Calibration (SDC). Unlike prior methods that regard model bias towards known categories as an obstacle to novel category identification, SDC provides a novel insight into unleashing the potential of the bias to facilitate novel category learning. Specifically, the output of the biased model serves two key purposes. First, it provides an accurate modeling of category bias, which can be utilized to measure the degree of bias and debias the output of the current training model. Second, it offers valuable insights for distinguishing different novel categories by transferring knowledge between similar categories. Based on these insights, SDC dynamically adjusts the output logits of the current training model using the output of the biased model. This approach produces less biased logits to effectively address the issue of category bias towards known categories, and generates more accurate pseudo labels for unlabeled data, thereby mitigating category confusion for novel categories. Experiments on three benchmark datasets show that SDC outperforms SOTA methods, especially in the identification of novel categories. Our code and data are available at \url{https://github.com/Lackel/SDC}.

Figures (5)

• Figure 1: Four types of GCD errors. For instance, Novel-Novel Error indicates samples from novel categories are misclassified into wrong novel categories.
• Figure 2: The overall framework of our model, which mainly contains Bias Mitigation, Confusion Mitigation, and Entropy-based Weighting. $\ominus$ means subtracting the biased logits from the original ones for known categories, while $\oplus$ means adding the transfer logits to the original ones for novel categories. $\hat{y}$ is the generated pseudo label utilizing the calibrated logits through the Sinkhorn-Knopp (S-K) algorithm.
• Figure 3: Entropy distribution for known and Novel categories based on the biased logits.
• Figure 4: Accuracy of pseudo labels for all (All) and novel (Novel) categories with and without logit adjustment (LA).
• Figure 5: Visualization of the learned features and model performance under different known category ratios (KCR).