Data augmentation: Artificially multiplying training data

In short: In data augmentation existing training data is changed slightly (rotated, cropped, noised) so that the model becomes more robust and overfits less.

In more detail: Good AI needs much and varied data. If it is scarce, several changed variants are produced from each example without the correct answer changing: a mirrored cat remains a cat.

In Depth

Images

Typical changes: flipping, rotating, cropping and scaling, brightness/contrast/hue, blur and noise, shifting, cutout (hiding random areas), mixup/CutMix (mixing images and labels). Libraries: torchvision.transforms, Albumentations, Keras layers. Important: only plausible changes, a digit “6” rotated becomes a “9”.

Text and audio

Text: replace words with synonyms, rearrange sentences, back-translation (German → English → German), random deletion or swapping, generating variants with a language model. Audio: change tempo and pitch, mix in background noise, SpecAugment (masking time/frequency regions).

Tables and synthetic data

With tables it is delicate; methods such as SMOTE generate additional examples of rare classes. Simulation and generative models produce synthetic data (for example vehicles in dangerous situations). Check whether synthetic data match reality and do not amplify biases. With personal data synthesis can help but does not replace a privacy check.

Practice

Augment only the training set, never validation and test. Random changes are drawn anew every epoch, so the model never sees exactly the same thing twice. Watch whether validation improves; too strong augmentation hurts. It is a cheap way against overfitting but does not replace real, varied data.

See also: Regularisation and early stopping, Convolutional networks (CNN), Bias, fairness and data quality, Training, validation and test