ProtoGCD: Unified and Unbiased Prototype Learning for Generalized Category Discovery.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Generalized Category Discovery (GCD) aims to cluster and identify novel categories using labeled data from known classes.
• Existing methods struggle with imbalanced accuracy and biased representations due to separate handling of old/new classes or decoupled objectives.

Purpose of the Study:

• To develop a unified and unbiased prototype learning framework for Generalized Category Discovery.
• To improve representation learning and clustering accuracy for both known and novel categories.

Main Methods:

• Introduced ProtoGCD, a unified prototype learning framework with joint prototypes and learning objectives.
• Proposed a dual-level adaptive pseudo-labeling mechanism to reduce confirmation bias.
• Incorporated regularization terms for enhanced representation learning and devised a criterion for estimating the number of new classes.

Main Results:

• ProtoGCD achieves state-of-the-art performance on both generic and fine-grained datasets.
• The framework enables unified modeling of old and new classes, mitigating accuracy imbalances.
• Extended ProtoGCD effectively detects unseen outliers, demonstrating task-level unification.

Conclusions:

• ProtoGCD offers a significant advancement in Generalized Category Discovery by unifying class modeling.
• The proposed methods effectively address limitations of prior approaches, leading to superior performance.
• The framework's ability to detect outliers further enhances its practical applicability.