FixMatch is a semi-supervised learning algorithm that combines pseudo-labeling with consistency regularization — using weak augmentation to generate confident pseudo-labels and strong augmentation to create challenging training targets, achieving near-supervised performance with very few labels.

How Does FixMatch Work?

• Weak Augmentation: Apply weak augmentation (flip, crop) to unlabeled data -> generate prediction.
• Pseudo-Label: If $max(p_{weak}) > au$ (typically $ au = 0.95$), use $argmax(p_{weak})$ as a hard pseudo-label.
• Strong Augmentation: Apply strong augmentation (RandAugment, CTAugment) to the same unlabeled image.
• Loss: Cross-entropy between the pseudo-label and the model's prediction on the strongly augmented version.
• Paper: Sohn et al. (2020).

Why It Matters

• Simplicity: Two simple ideas (confidence pseudo-labeling + weak/strong augmentation) combined elegantly.
• Few Labels: 250 labels on CIFAR-10 → 94.9% accuracy (vs. 95.0% supervised with 50K labels).
• Standard: Became the baseline for semi-supervised learning research.

FixMatch is the elegant union of pseudo-labeling and consistency — using weak views for labels and strong views for training in a remarkably effective combination.