Definition
Data augmentation creates additional training examples by applying transformations that should preserve the target label or another task-relevant relationship. Common image transformations include crops, reflections, color changes, and rotations. Audio can be shifted or mixed with noise. Text augmentation may paraphrase an example or alter a controlled attribute, but meaning is easier to damage than pixels.
An augmentation policy is an encoded claim about invariance. A horizontal reflection may preserve the label for a dog photograph but reverse the meaning of handwriting, road signs, or medical anatomy. Good evaluation therefore checks the transformed examples, subgroup effects, and performance on untouched validation data rather than assuming that a larger training set is better.
An influential ImageNet example
Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton documented an influential large-scale use in their 2012 ImageNet paper. They generated random crops and horizontal reflections, then perturbed RGB intensities. The paper reported these methods as ways to reduce overfitting in a 60-million-parameter convolutional network; it did not coin the general practice.
Distinguish it from nearby terms
Synthetic data can be generated independently from a simulator or generative model. Augmentation usually starts from an existing example or a task-specific rule for producing a valid variant. Regularization is broader and includes methods such as weight decay or dropout that do not create examples.
Check your understanding
You augment photographs of digits by rotating every image 180 degrees while keeping the original label. Which labels may change meaning, and what validation would catch the error before training?