Prototype-Neighbor Networks with task-specific enhanced meta-learning for few-shot classification.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Few-Shot Classification (FSC) relies on Prototypical Networks (PN), but their unimodal prototypes may not capture complex data distributions.
• Existing methods struggle with limited labeled data, impacting model representativeness.

Purpose of the Study:

• To introduce Prototype-Neighbor Networks (PNN) for improved Few-Shot Classification.
• To enhance the meta-learning mechanism for better adaptability to new classes.
• To develop a novel data augmentation method for FSC.

Main Methods:

• Propose Neighbor Network (NN) to classify samples based on neighbors and optimize the metric space.
• Combine PN and NN into PNN for robust metric learning with limited data.
• Incorporate task-specific fine-tuning and a PN-NN data augmentation technique to reduce pseudo-label noise.

Main Results:

• PNN outperforms 24 state-of-the-art FSC algorithms on mini-imageNet and CUB datasets.
• Achieved competitive results on tiered-imageNet.
• Demonstrated effectiveness on four cross-domain medical image datasets.

Conclusions:

• PNN learns a superior metric space for Few-Shot Classification using limited data.
• The enhanced meta-learning and data augmentation improve model generalization and reduce noise.
• PNN shows significant potential for both general and cross-domain FSC tasks.