VPAF-FSCIL: Virtual prototype calibration and parameter-adaptive freezing for few-shot class-incremental learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Class Incremental Learning (CIL) models learn new image classes while retaining old ones.
• Conventional CIL requires abundant data, which is often unavailable in real-world scenarios like autonomous driving and medical diagnosis.
• Few-Shot Class Incremental Learning (FSCIL) addresses data scarcity but struggles with catastrophic forgetting, feature drift, and class imbalance.

Purpose of the Study:

• To propose a novel FSCIL method that improves feature stability and adaptability to new classes.
• To overcome the limitations of existing FSCIL methods, including catastrophic forgetting and feature drift.
• To enhance classification accuracy and model stability in few-shot incremental learning scenarios.

Main Methods:

• A feature-decoupled network architecture comprising a Feature Extractor, Feature Distributor, and ETF Classifier.
• Integration of adaptive virtual class space allocation for optimal new class embeddings.
• Parameter-Adaptive Freezing (PAF) strategy utilizing the Fisher Information Matrix to selectively freeze parameters.

Main Results:

• The proposed method demonstrates superior performance compared to state-of-the-art FSCIL techniques.
• Significant improvements in classification accuracy and model stability were observed across multiple benchmark datasets (CIFAR-100, miniImageNet, CUB-200-2011).
• Effective mitigation of catastrophic forgetting and feature drift in few-shot incremental learning.

Conclusions:

• The developed FSCIL method effectively balances learning new classes with retaining knowledge of old classes.
• The feature-decoupled architecture and PAF strategy provide a robust solution for data-scarce incremental learning.
• This approach offers a promising direction for real-world applications requiring continuous learning from limited data.