Progressive Learning Strategy for Few-Shot Class-Incremental Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Few-shot class incremental learning (FSCIL) aims to learn new concepts with limited data while retaining prior knowledge.
• Current FSCIL methods often freeze feature extractors after base training, leading to overfitting and forgetting when new data is introduced.
• Overfitting on challenging samples degrades decision boundary robustness and exacerbates knowledge forgetting in incremental learning.

Purpose of the Study:

• To propose a progressive learning strategy (PGLS) to enhance robustness and generalization in FSCIL.
• To address overfitting and knowledge forgetting issues in traditional FSCIL frameworks.
• To improve the adaptability of base classes to novel classes in incremental learning scenarios.

Main Methods:

• Developed a covariance noise perturbation approach for sample robustness assessment, inspired by curriculum learning.
• Implemented a sample reweighting scheme prioritizing robust samples for stability, then weakly robust samples for generalization.
• Introduced a curriculum learning strategy with progressive virtual class augmentation during base training for forward compatibility.

Main Results:

• The proposed PGLS method demonstrated significant advantages over state-of-the-art models.
• Experiments on CUB200, CIFAR100, and miniImageNet datasets validated the effectiveness of the approach.
• The strategy successfully enhanced model adaptability and alleviated forgetting problems.

Conclusions:

• The PGLS strategy effectively improves robustness and generalization in few-shot class incremental learning.
• Progressive learning and sample reweighting are key to mitigating overfitting and forgetting.
• The proposed method offers a promising direction for advancing FSCIL research.