AREA: Attribute Extraction and Aggregation for CLIP-Based Class-Incremental Learning

Researchers introduced AREA, a new class-incremental learning method for CLIP models.

AREA splits recognition into attribute extraction and attribute aggregation, then stabilizes both with geometric and information-bottleneck techniques.

At inference time, it uses task routing to better handle new categories without overwriting old ones.

The method is designed to reduce catastrophic forgetting, a major challenge in continual CLIP learning.

The paper reports that AREA outperforms existing approaches in experiments.