A team training a model to spot diseased leaves from farmers’ phone photos may have only a few thousand labeled pictures. Augmentation shows the model each photo several ways during training: flipped, slightly rotated, cropped, darker or brighter. A diseased leaf is still diseased when the photo is tilted, so the label stays correct, and the model learns to ignore angle and lighting instead of memorizing them.
The edits have to respect the task. Rotating a handwritten 6 far enough turns it into a 9; swapping words in a product review can flip its sentiment. Some methods change labels on purpose: mixup blends two images and their labels in proportion. For the plant example, evaluate on held-out photos that resemble real use. Deliberate test-time augmentation is another option: run several crops or flips through the model and combine their predictions, as PyTorch’s TenCrop example does. Evaluate the same prediction procedure you intend to deploy, and keep evaluation examples separate from training.
Augmentation starts from real examples and perturbs them under rules you choose. Synthetic data is generated, often by another model, to stand in for data you don’t have. Both are cheaper than labeling more real examples, and neither replaces checking the model on real ones. Augmentation also has limits: it can’t teach the model about cases that none of the originals cover.
Sources
- TensorFlow: Data augmentation tutorial — Shows random flips and rotations applied to training images and notes that augmentation layers are inactive at test time, during evaluation and prediction; checked October 24, 2026.
- Zhang et al.: mixup: Beyond Empirical Risk Minimization — Proposes training on blended pairs of examples and their labels, an augmentation that deliberately combines examples instead of preserving one label; checked October 24, 2026.
- PyTorch torchvision: TenCrop — Shows test-time augmentation: evaluate crops and horizontal flips of an image and average the model's predictions; checked October 24, 2026.