Automatic Metric Search for Few-Shot Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Vision

Background:

• Few-shot learning (FSL) aims to classify new categories with minimal labeled data.
• Current FSL methods often rely on manually engineered metric functions, demanding significant expertise and effort.
• This reliance on predefined metrics limits adaptability and performance across diverse tasks.

Purpose of the Study:

• To develop a novel approach for automated metric function discovery in few-shot learning.
• To introduce the automatic metric search (Auto-MS) framework for task-specific metric optimization.
• To enhance the efficiency and effectiveness of few-shot learning models.

Main Methods:

• Proposed the automatic metric search (Auto-MS) model with a dedicated Auto-MS space.
• Developed a novel search strategy integrating episode-training with a bilevel optimization approach.
• Optimized both network weights and structural parameters of the few-shot model.

Main Results:

• The Auto-MS model demonstrated superior performance on benchmark datasets (miniImageNet and tieredImageNet).
• Automated metric search proved more effective than manually predefined metrics.
• The proposed search strategy successfully optimized few-shot model parameters.

Conclusions:

• Auto-MS offers a powerful, automated solution for metric learning in few-shot scenarios.
• The developed search strategy effectively enhances few-shot model generalization.
• This work paves the way for more adaptable and efficient few-shot learning systems.