Other meanings of Data augmentation
Machine learning
Data augmentation is a machine-learning technique that increases training-data diversity by applying label-preserving transformations or by synthesizing new examples. It can reduce overfitting, improve robustness to ordinary variation, and make limited datasets more useful, but poorly chosen transformations can alter the target label or encode unrealistic assumptions.
Data augmentation expands the effective variety of a training set without requiring a separately collected label for every transformed example. In image learning, a pipeline may randomly crop, flip, rotate, resize, blur, or adjust color before an image reaches the model. The transformation is useful only when the altered image should retain its original class or annotation. Augmentation therefore acts as an inductive bias: it tells the learner which changes are irrelevant to the task. Research on deep learning identified augmentation, together with large datasets and optimization advances, as a major contributor to performance improvements in visual recognition.1 Augmented samples are usually generated on the fly, so a finite dataset can expose a model to many stochastic variants without storing them all.
Augmentation methods depend on the structure of the data and the invariances a task permits. Audio systems may alter pitch, tempo, loudness, or add background noise; natural-language systems may use constrained word substitution, deletion, or paraphrase, though preserving meaning is difficult. In images, mixup forms interpolations between examples and their labels, while CutMix combines image regions and proportionally mixes labels.23 Policies can also be learned rather than hand-designed: AutoAugment searches among transformation operations and magnitudes for a task-specific policy.4 Synthetic generation with a generative model is another form, but it introduces risks of artifacts, memorization, and distributional bias.
Augmentation improves generalization only when its variations resemble meaningful variation at deployment. A horizontal flip is often sensible for animal recognition, but may change the label for text, traffic signs, medical laterality, or asymmetrical objects. Excessive distortion can create ambiguous or impossible examples, while weak augmentation may have little effect. Augmentation must be applied after a careful train-validation-test split; transforming examples before splitting can create near-duplicates and leak information across evaluation sets. Reported gains should therefore be tested with ablations, fixed evaluation data, and checks for subgroup performance. In safety-critical settings, domain experts may need to approve transformations, especially where small geometric or photometric changes affect diagnosis or measurement.
Some of the most consequential augmentation choices concern annotations rather than pixels. A crop can remove the object of interest, a rotation can invalidate bounding boxes, and a spatial transform for segmentation must be applied consistently to both the image and its mask. Augmentation can also improve robustness to nuisance factors while leaving a model dependent on hidden shortcuts, so it is not a substitute for representative data collection.5 Automated policy search may discover transformations that raise benchmark accuracy but perform poorly under real-world shifts. Conversely, augmentation can be used deliberately for rare classes, privacy-preserving training, or robustness testing. Its success is consequently task-specific: the central question is not how much synthetic data to create, but which transformations describe credible variation.
Scope: this entry uses “data augmentation” in its machine-learning sense, not augmentation of databases, software interfaces, or human capabilities.
Help improve the encyclopedia. Reports go straight to the site manager.