Brain-Inspired Fast- and Slow-Update Prompt Tuning for Few-Shot Class-Incremental Learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Few-shot class-incremental learning (FSCIL) requires learning new classes with limited data sequentially.
• Foundation models with prompt tuning offer strong generalization but struggle with the dynamic nature of FSCIL.
• Existing prompt tuning methods are optimized for static datasets, not sequential learning challenges.

Purpose of the Study:

• To develop a novel prompt tuning method inspired by the brain's complementary learning systems (CLSs) for FSCIL.
• To enhance the transferability and adaptability of foundation models in incremental learning scenarios.
• To address the limitations of current prompt tuning techniques in handling sequential data and catastrophic forgetting.

Main Methods:

• Proposed Fast-and Slow-update Prompt Tuning FSCIL (FSPT-FSCIL), a brain-inspired approach.
• Categorized prompts into fast-update (for new knowledge) and slow-update (for meta-knowledge) groups.
• Employed interactive meta-learning for training prompts to balance rapid learning and knowledge retention.

Main Results:

• Demonstrated the effectiveness of FSPT-FSCIL through experiments on multiple benchmark datasets.
• Showcased the superiority of the proposed method compared to existing approaches.
• Validated the ability of FSPT-FSCIL to mitigate catastrophic forgetting in FSCIL.

Conclusions:

• FSPT-FSCIL offers a promising brain-inspired solution for few-shot class-incremental learning.
• The fast-and-slow update mechanism effectively balances learning new information and retaining existing knowledge.
• The method significantly improves foundation model performance in dynamic, sequential learning environments.