MMOOZZGG.com

Glossary

What is data augmentation

In short

Data augmentation is a way to expand a training set with changed copies of existing examples: rotated photos, reworded texts and voice recordings with added noise.

Machine learning · data augmentation

MMOOZZGG.com editorsChecked: 1 min read

For a child

You have one photo of a cat. You can flip it, make it a bit darker, zoom in or cut off an edge, and the cat is still a cat. This way one picture becomes a whole stack. A computer is also shown such stacks, so that it knows a cat in any light and in any corner of the photo.

In plain words

A model needs many examples, and collecting and labeling them takes a long time. So you take what you already have and change it a little: mirror and rotate a photo, change the brightness, cut out a piece, mix street noise into a voice recording, say a phrase in other words. The label stays the same. The model sees more variants and memorizes specific images less. An example: a document check trained on straight scans gets confused by phone photos, but after rotations and blur in training it handles them too.

For business

The method is cheaper than collecting and labeling new examples, so it is used in almost every project with photos, video and sound: quality control on the line, document recognition and speech in a call center. It helps when examples are few or a camera in the workshop takes worse pictures than one in the lab. Know the limits in advance. The changes must be realistic: mirrored text or an upside-down part will teach the model nonsense. The method adds no new facts: it will not create a defect that is not in any image. The costs here are mainly engineer time and computing during training. What gives a gain on your own data becomes clear in a pilot with an honest test set.

For a specialist

Data augmentation is the expansion of a training set with changed copies of examples, usually keeping the label. For images, teams use geometric changes (flip, rotation, crop, scale), photometric ones (brightness, contrast, color), erasing parts of the image (cutout, random erasing) and mixing examples, where the label changes too: mixup (Zhang and coauthors, 2017) trains a network on convex combinations of pairs of examples and their labels. For text, teams use synonym replacement, word insertion and swapping, and back-translation: Sennrich, Haddow and Birch (2015) translated monolingual texts in the target language back into the source language and added the resulting pairs to the parallel data. The survey by Feng and coauthors (2021) organizes the methods for NLP and notes that the discrete nature of language makes the task harder. Augmentation counts as regularization: it reduces overfitting but does not replace data from the right distribution. Policies of changes are also chosen automatically (AutoAugment, RandAugment). The test set is usually left unchanged, or the quality estimate is biased.

Example

A tile maker collected only a few hundred photos with chips. The engineers added rotations, glare and light blur to them, and the model began to find chips in images from a new camera mounted at a different angle.

Related terms

Questions and answers

How is data augmentation different from synthetic data? Augmentation changes examples you already have and keeps their label: the same image, only rotated. Synthetic data is created from scratch by a generator or a simulation.

Generative models blur the line, because they draw new variants based on real samples.

When does augmentation do harm? When a change alters the meaning of an example: a mirrored letter “b” turns into “d”, and a flipped X-ray mixes up the left and right sides. Then the model learns from wrong labels.

So choose the set of changes for the task and check it on the test set.

Sources

Not sure whether your company needs this?

AI solutions map
Contact us