Remixmatch

ReMixMatch is a semi-supervised learning algorithm that extends MixMatch with distribution alignment and augmentation anchoring — using strong augmentations guided by a weakly augmented "anchor" to generate better training targets for unlabeled data.

Key Components of ReMixMatch

• Distribution Alignment: Adjust pseudo-label distribution to match the labeled data's class distribution.
• Augmentation Anchoring: Generate pseudo-labels from weakly augmented input, then train on multiple strongly augmented versions.
• CTAugment: Learned augmentation policy that adapts augmentation magnitude based on network confidence.
• Self-Supervised Rotation: Additional rotation prediction loss as auxiliary task.
• Paper: Berthelot et al. (2020).

Why It Matters

• Class Balance: Distribution alignment prevents the model from being biased toward majority pseudo-label classes.
• Better Than MixMatch: Significant accuracy improvement over MixMatch, especially with very few labels.
• Augmentation Bridge: Anchoring bridges weak and strong augmentations effectively.

ReMixMatch is MixMatch with class balance and augmentation control — adding distribution-aware pseudo-labeling and adaptive augmentation for better semi-supervised learning.