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: $\mathcal{L}_{UDA} = \text{KL}(p(y|x) p(y|\text{Aug}(x)))$ — 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.