Pseudo-Labeling Based Practical Semi-Supervised Meta-Training for Few-Shot Learning.

Area of Science:

• Machine Learning
• Computer Science

Background:

• Few-shot learning (FSL) methods often require extensive labeled data for meta-training, limiting their practical application.
• Existing semi-supervised meta-training (SSMT) approaches for FSL necessitate class-aware sample selection from unlabeled data, which is often impractical.

Purpose of the Study:

• To propose a practical semi-supervised meta-training setting for FSL using truly unlabeled data.
• To introduce a novel meta-training framework, pseudo-labeling based meta-learning (PLML), to effectively leverage both labeled and unlabeled data.

Main Methods:

• A classifier is trained using semi-supervised learning (SSL) to generate pseudo-labels for unlabeled data.
• Few-shot tasks are constructed using labeled and pseudo-labeled data.
• A novel fine-tuning method incorporating feature smoothing and noise suppression is employed to train FSL models on potentially noisy pseudo-labels.

Main Results:

• The proposed PLML framework effectively mitigates performance degradation in FSL models with limited labeled data.
• PLML significantly outperforms existing representative SSMT models on FSL benchmarks.
• The meta-training approach also enhances the performance of several SSL algorithms.

Conclusions:

• PLML offers a practical and effective solution for semi-supervised meta-training in few-shot learning, especially in scenarios with limited labeled data.
• The framework's ability to utilize truly unlabeled data makes it suitable for realistic applications.
• This approach demonstrates the potential of pseudo-labeling for improving both FSL and SSL methods.