FlexMatch is a semi-supervised learning algorithm that extends FixMatch with class-specific flexible confidence thresholds — allowing easy-to-learn classes to have higher thresholds and hard classes to have lower thresholds, improving learning fairness across classes.

How Does FlexMatch Work?

• Curriculum: Start with a lower threshold and increase it as the model improves.
• Per-Class: Each class has its own dynamic threshold based on its learning status.
• Learning Status: Track how well the model predicts each class on unlabeled data.
• Threshold: Classes that the model already handles well get higher thresholds. Struggling classes get lower thresholds.
• Paper: Zhang et al. (2021).

Why It Matters

• Class Fairness: FixMatch's fixed threshold causes the model to ignore hard classes early in training — FlexMatch fixes this.
• Curriculum Learning: The adaptive threshold naturally creates a curriculum from easy to hard classes.
• SOTA: Outperforms FixMatch significantly, especially with very few labels per class.

FlexMatch is FixMatch with class-adaptive confidence — ensuring every class gets a fair chance to contribute pseudo-labels during training.