Language-Inspired Relation Transfer for Few-Shot Class-Incremental Learning

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-Shot Class-Incremental Learning (FSCIL) aims to enable systems to learn new classes from limited examples while retaining existing knowledge.
• Current FSCIL methods often struggle with a trade-off between base and incremental knowledge due to reliance on visual encoder tuning.
• Human learning effectively incorporates language descriptions for recognizing novel concepts, a capability lacking in many AI systems.

Purpose of the Study:

• To propose a novel Language-inspired Relation Transfer (LRT) paradigm for Few-Shot Class-Incremental Learning (FSCIL).
• To leverage both visual cues and textual descriptions for improved object understanding and classification in open-world settings.
• To overcome the limitations of existing methods by addressing the knowledge trade-off and domain gap in incremental learning.

Main Methods:

• Developed a two-step LRT paradigm integrating visual and language information.
• Introduced a graph relation transformation module to transfer pre-trained text knowledge to visual domains.
• Implemented a text-vision prototypical fusion module for combining visual and language embeddings.
• Utilized context prompt learning for rapid domain alignment and imagined contrastive learning to address limited text data during alignment.

Main Results:

• The proposed LRT paradigm demonstrated superior performance compared to state-of-the-art models.
• Achieved over 13% improvement on the miniImageNet FSCIL benchmark.
• Achieved over 7% improvement on the CIFAR-100 FSCIL benchmark.

Conclusions:

• The LRT paradigm effectively enhances Few-Shot Class-Incremental Learning by synergizing visual and language modalities.
• The proposed methods for domain alignment and text-image transfer successfully mitigate challenges in incremental learning.
• LRT offers a promising direction for developing more robust and adaptable AI systems capable of lifelong learning.