Ensemble Transductive Propagation Network for Semi-Supervised Few-Shot Learning.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Few-shot learning faces challenges with limited training data, leading to high variance, bias, and overfitting.
• Graph-based transductive few-shot learning utilizes unlabeled data to improve predictions, emerging as a significant research area.
• Existing methods struggle to effectively integrate information from limited labeled and abundant unlabeled data.

Purpose of the Study:

• To propose a novel ensemble semi-supervised few-shot learning strategy named ensemble transductive propagation networks (ETPN).
• To enhance the utilization of unlabeled data and improve the stability and accuracy of few-shot learning models.
• To address the limitations of current few-shot learning approaches by integrating transductive networks and Dempster-Shafer evidence fusion.

Main Methods:

• Developed homogeneity and heterogeneity ensemble transductive propagation networks with preset weight coefficients for iterative inference.
• Improved Dempster-Shafer (D-S) evidence fusion by incorporating information entropy for stable multi-model result fusion.
• Enhanced ensemble pruning using L2 norm to select accurate individual learners and introduced interference sets for improved anti-disturbance capabilities.

Main Results:

• The proposed ensemble transductive propagation networks (ETPN) demonstrated superior performance compared to state-of-the-art few-shot learning models.
• Achieved accuracy improvements of 0.3% and 0.28% in 5-way 5-shot settings on miniImagNet and tieredImageNet, respectively.
• Showcased significant gains of 3.43% and 7.6% in 5-way 1-shot settings on miniImagNet and tieredImageNet, highlighting effectiveness with extremely limited data.

Conclusions:

• ETPN effectively addresses the challenges of few-shot learning by robustly utilizing unlabeled data through transductive learning and evidence fusion.
• The novel integration of ensemble methods, D-S theory, and transductive propagation offers a stable and accurate approach for few-shot classification.
• The proposed strategy represents a significant advancement in few-shot learning, particularly in scenarios with scarce labeled samples.