DyCR: A Dynamic Clustering and Recovering Network for Few-Shot Class-Incremental Learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Few-shot class-incremental learning (FSCIL) faces catastrophic forgetting, where models lose previously learned information when trained on new data.
• Existing static methods struggle with adapting old knowledge to new data, limiting performance in continual learning scenarios.

Purpose of the Study:

• To propose a dynamic clustering and recovering network (DyCR) that addresses the adaptation problem and mitigates forgetting in FSCIL.
• To enable trainable and dynamic networks that learn new features and adapt better to novel data during incremental stages.

Main Methods:

• Developed a dynamic and trainable network (DyCR) for FSCIL, contrasting with static approaches.
• Introduced an orthogonal decomposition mechanism to split feature embeddings into context and category information.
• Utilized preserved context information for recovering old class features and optimized the feature embedding space using category information.

Main Results:

• The DyCR network demonstrated superior performance compared to existing methods across four benchmark datasets.
• The orthogonal decomposition effectively mitigated catastrophic forgetting by preserving and recovering old class features with reduced data requirements.
• The category-aware optimization enhanced feature discrimination and intra-class compactness.

Conclusions:

• The proposed DyCR network offers an effective solution for few-shot class-incremental learning by dynamically adapting to new data and preserving old knowledge.
• The novel orthogonal decomposition mechanism provides a memory-efficient way to combat catastrophic forgetting.
• DyCR advances the state-of-the-art in continual learning for scenarios with limited data per class.