uda

**UDA** (Unsupervised Data Augmentation) is a **semi-supervised learning framework that applies advanced data augmentation (such as back-translation for text and RandAugment for images) to unlabeled data** — enforcing consistency between the original and augmented versions. **How Does UDA Work?** - **Strong Augmentation**: Apply task-specific strong augmentation to unlabeled data (back-translation, RandAugment, TF-IDF word replacement). - **Consistency**: $mathcal{L}_{UDA} = ext{KL}(p(y|x) || p(y| ext{Aug}(x)))$ — predictions should be consistent under augmentation. - **Confidence Masking**: Only compute the consistency loss when the model is confident about $p(y|x)$. - **Paper**: Xie et al. (2020, Google Brain). **Why It Matters** - **Text + Vision**: One of the first methods to show strong semi-supervised results on both text (IMDb, BERT) and vision (CIFAR, ImageNet). - **Augmentation Is Key**: The quality of the augmentation strategy is the primary driver of performance. - **Low-Label**: 20 labels on IMDb → competitive with fully supervised BERT using 25K labels. **UDA** is **augmentation-powered semi-supervised learning** — leveraging the best task-specific augmentations to extract maximum value from unlabeled data.