Prototype Completion for Few-Shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot learning (FSL) aims to classify new categories using minimal data.
• Current pre-training methods fine-tune feature extractors, yielding limited gains due to class distribution issues in feature space.
• Base classes form compact clusters, while novel classes exhibit high variance, making feature extractor fine-tuning less effective.

Purpose of the Study:

• To address the limitations of fine-tuning in few-shot learning.
• To propose a novel meta-learning framework focused on improving prototype representation.
• To enhance the accuracy and effectiveness of few-shot classification models.

Main Methods:

• A prototype completion meta-learning framework is introduced.
• Utilizes primitive knowledge (part/attribute annotations) to extract prior features for seen attributes.
• Employs a part/attribute transfer network to infer features for unseen attributes.
• A prototype completion network generates enhanced prototypes.
• A Gaussian-based prototype fusion strategy leverages unlabeled samples to combine mean-based and completed prototypes.
• An economic version without primitive knowledge is also developed for fair comparison.

Main Results:

• The proposed method generates more accurate prototypes compared to existing techniques.
• Achieves superior performance in both inductive and transductive few-shot learning settings.
• Demonstrates the effectiveness of prototype completion and fusion strategies.

Conclusions:

• Fine-tuning feature extractors is suboptimal for few-shot learning due to inherent data distributions.
• Estimating representative prototypes through completion and fusion is a more effective strategy.
• The developed framework significantly advances few-shot learning capabilities.