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.