Data Augmentation

Data Augmentation artificially expands a training dataset by applying realistic, label-preserving transformations to existing data, boosting model generalization and mitigating overfitting.

This technique is a critical component in deep learning pipelines, especially when training data is limited: it generates new, synthetic samples from your original set. For image data, operations like random rotation, horizontal flipping, or color jittering create diverse perspectives, effectively multiplying the dataset size (e.g., one image becomes four or more). In Natural Language Processing (NLP), augmentation involves techniques like synonym replacement or random word insertion/deletion. The goal is to improve model robustness, ensuring the final algorithm (like a Convolutional Neural Network or CNN) performs accurately on unseen, real-world variations, which directly reduces the risk of catastrophic overfitting.

https://www.tensorflow.org/guide/data_augmentation