Transductive Relation-Propagation With Decoupling Training for Few-Shot Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Few-shot learning (FSL) aims to classify new concepts using minimal labeled data.
• Existing FSL methods struggle with task adaptation and leveraging support-query pair relations.
• This limits their performance on novel tasks and real-world applications.

Purpose of the Study:

• To develop a novel few-shot learning model that addresses limitations of existing methods.
• To improve task adaptability and performance by explicitly modeling relations between support and query samples.
• To introduce a decoupling training strategy for enhanced knowledge transfer in few-shot learning.

Main Methods:

• Proposed a transductive relation-propagation graph neural network (TRPN) for few-shot learning.
• Introduced a decoupling training strategy (TRPN-D) to separate embedding network and graph module training.
• Modeled support-query pair relations as graph nodes (relational nodes) for relation propagation.

Main Results:

• The TRPN-D model demonstrated significant performance improvements over state-of-the-art few-shot learning methods.
• Relation propagation effectively generated discriminative relation embeddings for support-query pairs.
• The decoupling strategy enabled effective knowledge transfer to new tasks.

Conclusions:

• The proposed TRPN-D method offers a novel and effective approach to few-shot learning.
• Explicitly modeling and propagating relations significantly enhances FSL performance.
• Decoupling training provides a new paradigm for improving knowledge transfer in FSL.