Uncertainty-weighted semi-supervised learning with dynamic entropy masking and Bhattacharyya-regularized loss

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Semi-supervised learning (SSL) utilizes labeled and unlabeled data for classification.
• Current SSL methods struggle with noisy, class-imbalanced datasets due to underutilizing uncertain samples and propagating pseudo-label errors.

Purpose of the Study:

• To develop an SSL framework that enhances performance on challenging datasets by effectively managing sample uncertainty.
• To improve model generalization and robustness in low-data, noisy, and imbalanced scenarios.

Main Methods:

• Introduced an uncertainty-weighted training mechanism prioritizing moderately uncertain samples.
• Implemented a dynamic entropy mask to defer extremely uncertain samples, limiting error propagation.
• Combined masked cross-entropy with a Bhattacharyya-regularized alignment term for improved view consistency and distribution alignment.

Main Results:

• Achieved 3-5% absolute accuracy gains over strong SSL baselines (FixMatch, ReMixMatch, FreeMatch) on benchmark datasets (CIFAR-10, SVHN, STL-10).
• Demonstrated significant improvements in label-scarce and class-imbalanced settings.
• Showcased enhanced model generalization and robustness against label noise and imbalance.

Conclusions:

• The proposed SSL framework effectively addresses limitations of existing methods in handling data uncertainty.
• The dynamic entropy masking and uncertainty weighting significantly boost performance in challenging real-world scenarios.
• The approach offers a robust solution for semi-supervised classification with limited or noisy data.