ClearGCD: Mitigating Shortcut Learning For Robust Generalized Category Discovery

2025-11-28

AI Summary

Open-world generalized category discovery (GCD) suffers from prototype confusion caused by shortcut learning, leading to catastrophic forgetting of known classes and degraded discriminability for novel classes. To address this, we propose a semantic consistency enhancement and dynamic prototype alignment framework. Specifically, we design a semantic view alignment mechanism: weak augmentations preserve semantic stability, while strong augmentations—generated via cross-class image patch replacement—suppress non-semantic shortcuts. We further introduce shortcut-suppressing regularization and an adaptive prototype memory bank to support contrastive learning. Our method consistently outperforms state-of-the-art approaches across multiple benchmarks. It effectively retains knowledge of known classes while significantly improving separation of novel categories. The framework demonstrates strong generalization and robustness, and is plug-and-play compatible with various parametric GCD architectures.

Technology Category

• Machine Learning: Semi-Supervised Learning
• Computer Vision: Adversarial Attacks & Robustness
• Search and Optimization: Learning to Search

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs

Abstract

In open-world scenarios, Generalized Category Discovery (GCD) requires identifying both known and novel categories within unlabeled data. However, existing methods often suffer from prototype confusion caused by shortcut learning, which undermines generalization and leads to forgetting of known classes. We propose ClearGCD, a framework designed to mitigate reliance on non-semantic cues through two complementary mechanisms. First, Semantic View Alignment (SVA) generates strong augmentations via cross-class patch replacement and enforces semantic consistency using weak augmentations. Second, Shortcut Suppression Regularization (SSR) maintains an adaptive prototype bank that aligns known classes while encouraging separation of potential novel ones. ClearGCD can be seamlessly integrated into parametric GCD approaches and consistently outperforms state-of-the-art methods across multiple benchmarks.

Problem

• Mitigates shortcut learning in category discovery
• Reduces prototype confusion for better generalization
• Enhances identification of known and novel classes

Innovation

• Cross-class patch replacement for semantic view alignment
• Adaptive prototype bank for shortcut suppression regularization
• Seamless integration into parametric GCD frameworks