ClearGCD: Mitigating Shortcut Learning For Robust Generalized Category Discovery

TL;DR

ClearGCD introduces a novel framework that reduces shortcut learning in generalized category discovery, improving robustness and accuracy in identifying both known and novel categories in open-world data.

Contribution

It proposes two new mechanisms, Semantic View Alignment and Shortcut Suppression Regularization, to mitigate shortcut learning and enhance GCD performance.

Findings

Outperforms state-of-the-art methods on multiple benchmarks.

Effectively reduces prototype confusion caused by shortcut learning.

Enhances generalization to novel categories.

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.