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

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Continual learning requires models to adapt to new information without forgetting past knowledge.
• Generalized Category Discovery (GCD) aims to identify novel classes in unlabeled data.
• Continual Generalized Category Discovery (C-GCD) combines these challenges, focusing on incremental new class discovery without prior data storage.

Purpose of the Study:

• To address the underexplored C-GCD task in a practical setting with numerous new classes and extended periods.
• To overcome the conflicting objectives of discovering new classes and preventing catastrophic forgetting of old classes.
• To mitigate prediction bias (new classes mistaken for old) and hardness bias (severe forgetting of difficult old classes).

Main Methods:

• Introduced a debiased learning framework named Happy.
• Employed clustering-guided initialization for robust feature extraction.
• Utilized soft entropy regularization to enhance new class discovery and clustering.
• Implemented hardness-aware prototype sampling to reduce forgetting of previous classes.
• Extended Happy to Happy++ with advanced training strategies for stability and generalizability.
• Developed Happy++O for unified open-world classification, including out-of-distribution detection.

Main Results:

• The Happy framework effectively balances the discovery of new classes and the retention of old ones.
• Clustering-guided initialization and soft entropy regularization improved new class identification.
• Hardness-aware prototype sampling significantly reduced forgetting of previously learned classes.
• Happy++ and Happy++O demonstrated stable, generalizable, and proficient performance across various C-GCD scenarios.
• The methods achieved remarkable performance in managing C-GCD conflicts.

Conclusions:

• The proposed Happy framework offers a robust solution for the C-GCD problem, tackling key biases.
• The extended versions, Happy++ and Happy++O, provide enhanced capabilities for incremental learning and open-world classification.
• The research advances the field of continual learning by enabling more effective and practical class discovery in dynamic environments.