Happy++: Towards Stable and Unified Continual Generalized Category Discovery

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Continual learning is essential for adapting to evolving environments.
• Generalized Category Discovery (GCD) aims to discover new classes from unlabeled data.
• Continual Generalized Category Discovery (C-GCD) incrementally discovers new classes without forgetting old ones.

Purpose of the Study:

• To address the challenges of C-GCD in a practical setting with more new classes and longer periods.
• To mitigate prediction bias (new classes mistaken for old) and hardness bias (severe forgetting of difficult old classes).
• To develop a robust framework for incremental class discovery and knowledge retention.

Main Methods:

• Introduced Happy, a debiased learning framework with hardness-aware prototype sampling and soft entropy regularization.
• Employed clustering-guided initialization for robust feature extraction.
• Proposed soft entropy regularization to enhance new class clustering and hardness-aware prototype sampling to reduce forgetting.

Main Results:

• Happy effectively addresses prediction and hardness biases in C-GCD.
• Happy++ (enhanced Happy) demonstrates more stable and generalizable performance.
• Happy++O extends capabilities to detect out-of-distribution samples, serving as a unified open-world classifier.

Conclusions:

• The proposed Happy framework significantly improves performance in C-GCD tasks.
• The methods effectively balance the conflicting objectives of discovering new classes and preventing catastrophic forgetting.
• The extended Happy++O offers a unified solution for open-world classification in practical scenarios.