Learning forward-compatible and domain-invariant representations for cross-domain few-shot class-incremental...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Few-Shot Class-Incremental Learning (FSCIL) typically assumes data from a single domain.
• Real-world applications often involve incremental classes from different domains than base classes.
• Existing FSCIL methods struggle with domain shift and data scarcity.

Purpose of the Study:

• Introduce Cross-Domain Few-Shot Class-Incremental Learning (CDFSCIL) as a more realistic challenge.
• Develop a unified framework (FCDI) to address CDFSCIL.
• Learn representations compatible with future classes and robust to domain shifts.

Main Methods:

• Propose the Learning Forward-Compatible and Domain-Invariant Representations (FCDI) framework.
• Construct a structured representation space for stable incremental learning.
• Disentangle semantic and domain features for domain-invariant representations.
• Employ diverse domain variations to enhance robustness.

Main Results:

• FCDI significantly outperforms state-of-the-art approaches on the proposed CDFSCIL benchmark.
• The method demonstrates strong performance on standard FSCIL tasks.
• Experimental results validate the effectiveness of the FCDI framework.

Conclusions:

• CDFSCIL presents a more challenging and realistic scenario for incremental learning.
• The FCDI framework effectively addresses both data scarcity and domain shift in incremental learning.
• FCDI learns robust, domain-invariant representations for improved few-shot class-incremental learning.