Understanding Episode Hardness in Few-Shot Learning.

Area of Science:

• Machine Learning
• Computer Science

Background:

• Deep learning generalization is limited by scarce annotated data.
• Few-shot learning uses sampled tasks (episodes) to improve generalization.
• Existing methods for episode difficulty assessment are impractical.

Purpose of the Study:

• To analyze factors influencing episode hardness in few-shot learning.
• To propose an efficient hardness assessment technique for episode sampling.
• To improve the generalizability of deep learning models.

Main Methods:

• Algebraic analysis of factors affecting episode hardness.
• Development of the Inverse-Fisher Discriminant Ratio (IFDR) for hardness assessment.
• Introduction of class-level (CL) and class-pair-level (CPL) sampling schemes.

Main Results:

• Episode hardness is significantly influenced by the classes within an episode.
• IFDR provides an efficient pre-sampling hardness assessment.
• CL and CPL sampling schemes reduce quantification cost and guarantee distribution.

Conclusions:

• The proposed IFDR method and sampling schemes effectively improve episode selection for few-shot learning.
• This approach enhances deep learning model generalizability in data-scarce scenarios.
• The findings offer a practical solution for optimizing few-shot learning episode design.