data augmentation

Data augmentation transforms existing training data to increase diversity without collecting new data. **Why it works**: More training examples, regularization effect, robustness to variations, addresses data scarcity. **NLP techniques**: **Paraphrasing**: Rephrase with LLM or back-translation. **Synonym replacement**: Swap words with synonyms. **Random insertion/deletion/swap**: Perturb text randomly. **EDA (Easy Data Augmentation)**: Combination of simple operations. **Back-translation**: Translate to another language and back. **Mixup**: Blend examples in embedding space. **Advanced techniques**: Adversarial examples, counterfactual augmentation, LLM-generated variations. **Vision techniques**: Rotation, cropping, color jitter, cutout, mixup, cutmix, AutoAugment. **Best practices**: Preserve labels (augmentation shouldn't change meaning), domain-appropriate transforms, validate on non-augmented test set. **Trade-offs**: Too aggressive augmentation creates noise, computational overhead, may not improve if data already sufficient. **Tools**: TextAttack, nlpaug, Albumentations (vision). Foundational technique for improving model robustness and generalization.