Data Augmentation
Techniques for artificially increasing the size and diversity of training data by creating modified versions of existing examples.
Image Augmentation
Flipping, rotating, cropping, color jittering, adding noise, and elastic deformations. These transformations create new training examples that teach the model invariance to these changes. Libraries like Albumentations and torchvision.transforms make this easy.
Text Augmentation
Synonym replacement, back-translation (translate to another language and back), random insertion/deletion, and paraphrasing with LLMs. Text augmentation is trickier because small changes can alter meaning.
Why It Works
Augmentation acts as a regularizer, reducing overfitting by exposing the model to more variation. It's especially valuable when labeled data is scarce. Modern self-supervised methods like contrastive learning rely heavily on augmentation strategies.