Task-Distributionally Robust Data-Free Meta-Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning

Background:

• Data-Free Meta-Learning (DFML) enables few-shot learning from pre-trained models without original data.
• Existing DFML methods lack comprehensive robustness analysis, crucial for real-world applications.

Purpose of the Study:

• To systematically investigate DFML robustness and identify failure modes.
• To propose a trustworthy DFML framework mitigating identified vulnerabilities.

Main Methods:

• Identified Task-Distribution Shift (TDS) and Task-Distribution Corruption (TDC) as key vulnerabilities.
• Developed a framework with synthetic task reconstruction, meta-learning with task memory interpolation, and automatic model selection.
• Utilized model inversion for synthetic task generation and replay strategies for knowledge retention.

Main Results:

• The proposed framework significantly enhances DFML robustness against TDS and TDC.
• Demonstrated effective filtering of untrustworthy models and prevention of catastrophic forgetting.
• Experimental validation across diverse datasets confirmed method superiority.

Conclusions:

• The novel DFML framework effectively addresses critical robustness issues.
• This work provides a more secure and reliable approach to data-free meta-learning.
• The findings are crucial for deploying DFML in complex, real-world scenarios.