Meta-learning for few-shot open task recognition

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Few-shot learning research typically uses fixed N-way K-shot settings.
• Real-world applications require models to adapt to unknown task configurations (open-task setting).
• This necessitates generalization to unseen structural combinations, not just interpolation.

Purpose of the Study:

• To address the limitations of fixed evaluation settings in few-shot learning.
• To introduce and evaluate a method for structural generalization in few-shot learning.
• To propose an open-task evaluation framework that better reflects real-world deployments.

Main Methods:

• Formalized three regimes for structural generalization: cross-way, cross-shot, and cross-way-cross-shot.
• Proposed Open-MAML, a meta-learning enhancement with dynamic classifier construction.
• Integrated inner-loop learning rate adaptation and AdaDropBlock regularizer for stability and robustness.

Main Results:

• Open-MAML demonstrated consistent performance improvements across within-domain and cross-domain evaluations.
• Achieved 1-7% absolute accuracy gains under single-dimensional changes (cross-way/cross-shot).
• Achieved 3-6% absolute accuracy gains under two-dimensional changes (cross-way-cross-shot).

Conclusions:

• Open-task evaluation is crucial for studying structural generalization in few-shot learning.
• Open-MAML provides a robust and effective approach for few-shot learning in dynamic environments.
• The proposed framework offers a reproducible basis for future research in this area.