Recursive Confidence Training for Pseudo-Labeling Calibration in Semi-Supervised Few-Shot Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Semi-Supervised Few-Shot Learning (SSFSL) faces challenges with data scarcity.
• Classifiers trained on limited data often produce biased, inaccurate pseudo-labels for unlabeled data.
• Inaccurate pseudo-labels can negatively impact downstream learning tasks in SSFSL.

Purpose of the Study:

• To introduce a novel method, Certainty-Aware Recursive Confidence Training (CARCT), to improve pseudo-labeling accuracy in SSFSL.
• To develop a technique for selecting more informative pseudo-labeled data for classifier retraining.
• To enhance the generalization capability of classifiers in data-scarce environments.

Main Methods:

• CARCT utilizes confidence levels of pseudo-labels to identify informative data for retraining.
• A joint double-Gaussian model is employed to learn a semi-supervised Prior Confidence Distribution (ssPCD).
• ssPCD guides the selection of high- and low-confidence pseudo-labeled data for recursive training and pseudo-labeling calibration.

Main Results:

• CARCT demonstrates superior performance compared to state-of-the-art methods in extensive SSFSL experiments.
• The method effectively distinguishes between high- and low-confidence pseudo-labels using learned confidence distributions.
• Recursive confidence training leads to accurate pseudo-labeling of unlabeled data, enhancing classifier generalization.

Conclusions:

• CARCT successfully addresses the issue of biased pseudo-labels in SSFSL through confidence-aware retraining.
• The proposed ssPCD effectively aids in pseudo-labeling calibration, improving classifier performance.
• The recursive training mechanism and self-training aspect of CARCT offer a robust solution for data-scarce learning scenarios.