Embedding Space Allocation with Angle-Norm Joint Classifiers for few-shot class-incremental learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Few-shot class-incremental learning (FSCIL) enables agents to learn new classes with limited data while retaining old knowledge.
• Existing methods struggle with feature space occupation by current classes and insufficient samples for training.
• Virtual class methods and Nearest Class Mean (NCM) classifiers have limitations in handling new class alignment and sample imbalance.

Purpose of the Study:

• To propose a novel learning framework, SAAN (Space Allocation with Angle-Norm joint classifiers), to address FSCIL challenges.
• To provide a balanced feature space allocation for all classes and enhance classification criteria using norm differences.
• To improve the adaptability of intelligent agents in dynamic environments.

Main Methods:

• SAAN divides the feature space into dedicated subspaces for each learning session, guided by pre-set category centers.
• It establishes a norm distribution for each class to generate angle-norm joint logits, addressing sample imbalance.
• The framework integrates class-center guidance for space allocation and angle-norm joint classifiers.

Main Results:

• SAAN achieves state-of-the-art performance in few-shot class-incremental learning tasks.
• The proposed method effectively balances feature space allocation across classes.
• SAAN demonstrates significant improvements in classification accuracy, especially under sample imbalance.

Conclusions:

• The SAAN framework offers a robust solution for few-shot class-incremental learning.
• It can be seamlessly integrated as a plug-in module to enhance existing state-of-the-art methods.
• SAAN advances the capability of intelligent agents to adapt to evolving data distributions.