Dynamic Support Network for Few-Shot Class Incremental Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot class-incremental learning (FSCIL) faces challenges with catastrophic forgetting and overfitting.
• These issues stem from feature distribution crumbling in fixed feature spaces, causing class confusion.

Purpose of the Study:

• To propose a novel Dynamic Support Network (DSN) to address catastrophic forgetting and overfitting in FSCIL.
• To enhance feature representation capacity and mitigate class confusion during incremental learning.

Main Methods:

• DSN adaptively updates the network with compressive node expansion for incremental classes.
• It dynamically compresses the network via node self-activation to reduce overfitting.
• DSN selectively recalls old class distributions to support feature distributions and prevent class confusion.

Main Results:

• DSN significantly improves performance over baseline methods in FSCIL.
• The proposed method achieves new state-of-the-art results on CUB, CIFAR-100, and miniImage datasets.
• DSN effectively alleviates catastrophic forgetting and overfitting.

Conclusions:

• DSN offers a systematic solution for catastrophic forgetting and overfitting in FSCIL.
• The compressive node expansion and class distribution recalling mechanisms are key to DSN's success.
• DSN demonstrates strong potential for real-world incremental learning applications.