data augmentation

**Data Augmentation** — artificially expanding the training dataset by applying random transformations, improving generalization without collecting more data. **Common Techniques (Vision)** - **Geometric**: Random crop, flip, rotation, scaling, affine transforms - **Color**: Brightness, contrast, saturation, hue jitter - **Erasing**: Random erasing, Cutout (mask random patches) - **Mixing**: Mixup (blend two images + labels), CutMix (paste patches between images) - **Auto**: AutoAugment, RandAugment — learned or random augmentation policies **NLP Augmentation** - Synonym replacement, random insertion/deletion - Back-translation (translate to another language and back) - Token masking (MLM-style) **Key Principles** - Augmentations should preserve the label (flipping a cat is still a cat) - Stronger augmentation = more regularization but can hurt if too aggressive - Test-Time Augmentation (TTA): Average predictions over augmented copies at inference for a small accuracy boost **Data augmentation** is one of the simplest and most effective regularization techniques in deep learning.