Data Augmentation: Definition and Examples

Data augmentation is a fundamental strategy in machine learning that multiplies the size and diversity of a dataset without having to collect new real data. It relies on applying controlled transformations to existing data to generate new plausible training examples.

In computer vision, this can include rotations, crops, horizontal flips, brightness adjustments, or adding noise to images. In natural language processing (NLP), techniques include paraphrase, synonym substitution, back-translation, or synthetic text generation using large language models.

The main goal is to reduce overfitting by exposing the model to a greater variety of situations during training. A model trained on augmented data learns to be more robust to variations it will encounter in production, thus improving its generalization ability.

In prompt engineering, the concept of data augmentation applies indirectly: one can use an LLM to generate prompt variations, create synthetic datasets for fine-tuning, or produce diverse examples for few-shot learning. This approach is particularly useful when annotated data is scarce or expensive to obtain.