Few Shot Class Incremental Learning via Efficient Prototype Replay and Calibration.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Few-shot class incremental learning (FSCIL) presents significant challenges in real-world applications.
• Key issues include catastrophic forgetting of previously learned knowledge and overfitting to new categories with limited data.

Purpose of the Study:

• To propose an efficient prototype replay and calibration (EPRC) method to enhance FSCIL performance.
• To address both catastrophic forgetting and overfitting in incremental learning scenarios.

Main Methods:

• Effective pre-training using rotation and mix-up augmentations for a robust backbone.
• Meta-training with pseudo few-shot tasks to improve feature extractor and projection layer generalization.
• Incorporation of an even nonlinear transformation for prototype calibration and reduced inter-category correlation.
• Prototype replay and explicit regularization within the loss function to combat forgetting and enhance discriminability.

Main Results:

• The proposed EPRC method significantly improves classification performance.
• Demonstrated effectiveness on CIFAR-100 and miniImageNet datasets.
• Outperforms existing mainstream FSCIL methods.

Conclusions:

• The EPRC method offers an effective solution for few-shot class incremental learning.
• It successfully balances the retention of old knowledge with the learning of new categories.
• The approach shows strong potential for real-world incremental learning applications.