Boosting Semi-Supervised Learning with Entropy-Guided Adaptive Reward Maximization

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Semi-supervised learning (SSL) methods commonly use pseudo-labeling and consistency regularization.
• Existing SSL approaches often overvalue high-confidence samples and neglect informative low-confidence ones, hindering generalization.

Purpose of the Study:

• Introduce EntropyMatch, an entropy-driven SSL framework to address the confidence-for-weighting issue in existing methods.
• Redefine sample importance using prediction entropy for more effective leveraging of unlabeled data.

Main Methods:

• EntropyMatch utilizes a bidirectional weighting strategy: upward exploitation for reliable samples and downward exploration for uncertain samples.
• An adaptive training mechanism adjusts focus from exploration to exploitation based on model maturity.

Main Results:

• Experiments on eight benchmarks across image classification, facial expression recognition, and human action recognition validate EntropyMatch.
• Achieved state-of-the-art results, demonstrating superior data efficiency and generalization.

Conclusions:

• EntropyMatch offers a robust and effective approach to semi-supervised learning.
• The framework shows significant improvements in data efficiency and model generalization across diverse tasks.