PNS-Based Regularization for Class-Incremental Learning Feature Expansion
A new method addresses catastrophic forgetting in Class-Incremental Learning (CIL) by using Probability of Necessity and Sufficiency (PNS) regularization. Current expansion-based methods freeze old features but suffer from collisions due to spurious correlations. Intra-task spurious correlations cause reliance on shortcut features, while inter-task correlations create semantic confusion between visually similar classes. The proposed approach extends PNS to guide feature expansion, aiming to reduce these collisions. The research is published on arXiv with ID 2603.09145.
Key facts
- Method addresses catastrophic forgetting in Class-Incremental Learning
- Uses Probability of Necessity and Sufficiency (PNS) regularization
- Current expansion-based methods freeze old features
- Spurious correlations cause feature collisions
- Intra-task correlations lead to shortcut feature reliance
- Inter-task correlations cause semantic confusion
- PNS definition extended to expansion-based CIL
- Published on arXiv with ID 2603.09145
Entities
Institutions
- arXiv