Transductive Parameter-Free Propagation Framework for Few-Shot Distribution Rectification.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Few-shot learning (FSL) is hindered by scarce labeled data for novel classes.
• Existing methods often train embedding functions using auxiliary base-class data, facing domain gap issues.
• Current embedding rectification techniques address distinct challenges in isolation, limiting broader applicability.

Purpose of the Study:

• To propose a general embedding rectification framework to enhance few-shot learning model performance.
• To address the limitations of isolated solutions by introducing a unified approach.
• To improve the reliability of pseudo-labeling for maximizing data distribution.

Main Methods:

• Introduced a distribution propagation (DisP) layer for task-level rectification, enhancing inter-class margins and intra-class aggregation.
• Developed a prototype propagation (ProtoP) layer for prototype-query level rectification, moving prototypes towards ideal class centers.
• Implemented a distribution-based pseudo-labeling method, pseudo-query upgrade (PseQUp), for reliable sample selection without confidence scores.

Main Results:

• The proposed framework demonstrates effective embedding rectification by maximizing actual data distribution.
• The distribution-based pseudo-labeling method (PseQUp) provides more reliable pseudo-labeling samples.
• Empirical experiments validate the applicable and plug-and-play nature of the proposed methods in transfer and meta-learning scenarios.

Conclusions:

• A general embedding rectification framework is essential for improving few-shot learning performance.
• The proposed DisP and ProtoP layers offer effective task-level and prototype-query level rectification.
• The PseQUp method enhances pseudo-labeling reliability, contributing to a more robust FSL framework.