Similarity-based prototype reconstruction and feature reorganization for non-exemplar class incremental learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning (DL) models face catastrophic forgetting when trained incrementally.
• Non-exemplar class incremental learning (NECIL) mitigates this by using class prototypes instead of original data.
• Effective utilization of these prototypes is crucial for preventing knowledge loss in NECIL.

Purpose of the Study:

• To propose a novel method for Non-exemplar Class Incremental Learning (NECIL) that addresses catastrophic forgetting.
• To enhance the utilization of memorized class prototypes for model stability and performance.
• To maintain decision boundaries of previous tasks during incremental updates.

Main Methods:

• Developed a similarity-based prototype reconstruction and feature reorganization (SPRR) method.
• Implemented a feature reorganization mechanism to continuously adapt prototypes to evolving feature spaces.
• Introduced a similarity-based prototype reconstruction to leverage new data for reconstructing old data features.
• Incorporated a knowledge integration strategy for classifier training to align with previous feature spaces.

Main Results:

• The SPRR method demonstrated effectiveness in Non-exemplar Class Incremental Learning (NECIL).
• Performance was validated on benchmark datasets: CIFAR-100, TinyImageNet, and ImageNet-Sub.
• Experimental results confirmed the method's ability to alleviate catastrophic forgetting.

Conclusions:

• The proposed SPRR method offers a robust solution for NECIL by effectively managing prototypes.
• The approach enhances model stability and performance in incremental learning scenarios.
• SPRR successfully mitigates catastrophic forgetting without requiring storage of original data samples.