Specialty Digest

Global English editionLearners’ Study Dashboard — in development
The Knowledge Library for All

Data Augmentation

nounAI Data and Infrastructurealso augmentation, image augmentation, text augmentation

In one line

Data augmentation is a set of techniques that enlarge and diversify a training data set by creating modified copies of existing examples, such as flipped or cropped images, to help models generalize better.

Data augmentation is a group of techniques that make a training data set larger and more varied by creating modified versions of the examples it already contains. A photograph of a dog can be flipped, rotated slightly, cropped or brightened, and each version still shows a dog. Training on these variants teaches a model that such changes do not alter the answer, which helps it handle new inputs and lowers the risk of overfitting. Augmentation is a standard part of preparing training data, especially when collecting and labeling new examples is expensive [1].

Why it works

A model with millions of parameters can memorize a small data set, which gives high accuracy on training examples and poor accuracy on new ones. Augmentation fights this in two ways. It supplies more varied examples, so there is less to memorize. It also encodes prior knowledge about the world: an object stays the same object if it moves a few pixels, a sentence keeps its meaning if a word is replaced by a synonym. The model is nudged toward learning features that are stable under these changes, which is the property called invariance.

Augmentation acts as a form of regularization, as do dropout and weight decay, and it often gives larger gains than adjusting the model itself. The 2012 AlexNet network, a landmark in deep learning, used random crops, horizontal reflections and color changes to enlarge its training set and reduce overfitting [2].

Image augmentation

Images are the classic case, because many transformations preserve what the picture shows.

  • Geometric changes. Flips, rotations, scaling, translation, cropping and shearing.
  • Color changes. Adjusting brightness, contrast, saturation and hue.
  • Noise and blur. Adding grain or blur to imitate poor cameras and motion.
  • Occlusion. Cutout and random erasing hide patches to make models rely on the whole object.
  • Mixing. Mixup blends two images and their labels, while CutMix pastes a patch from one image onto another. Mixup improved generalization on several benchmarks [3].
  • Automated policies. Methods such as AutoAugment and RandAugment search for the best combination of transformations.

Care is needed: flipping a photograph of the digit 6 vertically makes it look like a 9, and flipping a chest X-ray can swap left and right organs. Augmentations must respect the meaning of the labels. See computer vision.

Text augmentation

Text is harder because small edits can change meaning. Common methods include:

  • Synonym replacement and random insertion, swap or deletion of words.
  • Back-translation. Translate a sentence to another language and back to produce a paraphrase.
  • Paraphrasing with models. Large language models rewrite sentences or generate new examples in a given style.
  • Template filling. Insert different names, places or numbers into patterns.
  • Embedding-space noise. Perturb the numeric representation instead of the words.

See natural language processing and synthetic data.

Audio, tabular and other data

  • Audio. Time stretching, pitch shifting, added background noise, room echo and SpecAugment, which masks parts of the spectrogram, help speech recognition.
  • Tabular data. Oversampling of rare classes with SMOTE, which interpolates between neighbors, helps with imbalance.
  • Time series. Jittering, scaling and window slicing.
  • Graphs and 3D data. Dropping nodes or edges, rotating point clouds.
  • Reinforcement learning. Randomizing observations improves robustness; see reinforcement learning.

Offline and online augmentation

Augmentation can be done offline, creating and storing new files before training, or online, applying random transformations on the fly each time an example is loaded. Online augmentation saves storage and gives the model a fresh variation each epoch, and is the usual choice in deep learning.

Best practices

  • Augment only the training set. Validation and test sets should stay in original form, so scores reflect real data; see validation data and testing data.
  • Split first, then augment, so that copies of one example do not land in both training and test sets.
  • Choose transformations that match real variation the model will face.
  • Check samples visually to confirm labels still hold.
  • Tune the strength of augmentation, because too much can hurt.

Limitations

Augmentation cannot create genuinely new information. It cannot fix biased or unrepresentative data, and it may even reinforce errors. Extreme transformations can produce unrealistic examples. It also adds computing cost. When data is poor in kind, collecting more real data or using transfer learning may help more.

Augmentation for imbalanced and rare cases

Many real data sets are imbalanced: fraudulent transactions may be well under one percent of all records, and defects on a production line are rare by design. A model trained on such data tends to ignore the rare class. Augmentation helps by creating more variants of rare examples, for instance by applying many transformations to the few images of a defect, or by using SMOTE to interpolate new records between existing rare ones. It is often combined with class weighting and with careful choice of evaluation measures such as recall and precision, since plain accuracy hides failures on small classes.

Generative augmentation

A newer approach uses generative models to produce training examples. Image generators can create scenes with rare weather, lighting or objects. Language models can write extra customer messages for each intent. Simulators can render driving scenes with exact labels, which is valuable for robotics and autonomous vehicles. The main risks are that generated examples may be unrealistic or repeat the generator’s own errors, and that training repeatedly on model outputs can reduce diversity over time. Teams therefore mix synthetic and real data, check samples by eye and measure the effect on a held-out real test set.

Measuring whether augmentation helps

The test of any augmentation is performance on untouched validation data. Practitioners run ablation experiments, turning individual transformations on and off, to see which help. A transformation that raises validation accuracy is kept, and one that lowers it is dropped. Because gains depend on the data set and model, settings that work for one project may not transfer to another.

FAQs

What is data augmentation? It is creating modified copies of training examples, such as flipped or noised versions, to enlarge and diversify a data set.

Why is data augmentation used? To reduce overfitting and improve how well a model performs on new data, especially when data is limited.

Does augmentation increase the size of the dataset? Effectively yes, though online augmentation generates variations during training instead of storing them.

Can data augmentation be used on text? Yes, with methods like synonym replacement, back-translation and model-based paraphrasing, but it needs care to preserve meaning.

Should I augment test data? No. Keep test data unmodified so evaluation reflects real conditions.

More in AI Data and Infrastructure

Key facts

  • Data augmentation is widely used in image classification, where random crops, flips and color changes are applied to training images1.
  • The 2012 AlexNet network used image translations, reflections and color changes to enlarge its training set and reduce overfitting2.
  • The mixup technique trains on blended pairs of images and labels, and improved generalization on several benchmarks3.

Related concepts

Quick checkWhat is data augmentation?Show answer

Data augmentation is a family of techniques that expand and diversify a training data set by creating modified versions of existing examples, such as flipped, cropped or noised copies, so that a model generalizes better and overfits less.

Sources

  1. Connor Shorten and Taghi Khoshgoftaar, A Survey on Image Data Augmentation for Deep Learning, Journal of Big Data, https://doi.org/10.1186/s40537-019-0197-0
  2. Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton, ImageNet Classification with Deep Convolutional Neural Networks, Advances in Neural Information Processing Systems, https://papers.nips.cc/paper/2012/hash/c399862d3b9d6b76c8436e924a68c45b-Abstract.html
  3. Hongyi Zhang et al., mixup: Beyond Empirical Risk Minimization, https://arxiv.org/abs/1710.09412

Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue