Word Embedding Distribution Propagation Graph Network for Few-Shot Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Graph Neural Networks

Background:

• Few-shot learning (FSL) aims to enable models to learn from limited data, mimicking human generalization capabilities.
• Existing graph neural networks (GNNs) in FSL often overlook crucial semantic information, focusing solely on labeled sample transfer.
• This limitation hinders the performance of GNNs in scenarios with scarce training examples.

Purpose of the Study:

• To develop a novel graph neural network architecture for few-shot learning that incorporates class semantic information.
• To improve classification performance in few-shot learning by leveraging word embeddings and semantic relationships.
• To address the limitations of current GNNs in FSL by integrating semantic understanding into the learning process.

Main Methods:

• Proposed a Word Embedding Distribution Propagation Graph Network (WPGN) that embeds semantic information into GNNs for FSL.
• Integrated an attention mechanism and utilized Mahalanobis distance for class similarity calculation.
• Employed the Funnel ReLU (FReLU) activation function and updated both point and word embedding distribution graphs.

Main Results:

• The WPGN demonstrated significant accuracy improvements on standard FSL benchmarks compared to baseline models.
• Accuracy gains of 9.03%, 4.56%, and 4.15% were observed on 5-way-1-shot, 5-way-2-shot, and 5-way-5-shot tasks, respectively.
• The results highlight the effectiveness of incorporating semantic information into GNNs for few-shot classification.

Conclusions:

• The proposed WPGN effectively utilizes semantic information to enhance few-shot learning performance.
• Integrating attention mechanisms and appropriate similarity metrics further boosts the model's generalization capabilities.
• This approach offers a promising direction for advancing machine learning models that learn efficiently from minimal data.