AI glossary · Basic · Data engineering · Computer vision
Data augmentation
Data augmentation means creating more training examples by slightly changing or transforming training data you already have. It lets a model learn from a wider variety of cases without collecting new data.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Data, ops & platforms
- Fields
- Data engineering, Computer vision
- Added
- 2026-10-08
How it works
The 'correct answer' of the original data stays the same; only its surface changes. A cat photo flipped horizontally, slightly cropped or with changed brightness is still a cat, so one photo yields several training examples. For text, the same meaning can be rephrased, or back-translation can be used: translating into another language and then back again. For speech, background noise is mixed in or the speaking speed is changed. Recently it has also become common to ask an LLM to write new similar examples, which overlaps with synthetic data.
Augmentation methods by data type
- Data augmentation
- Images
- Horizontal flip
- Cropping
- Brightness change
- Text
- Paraphrasing
- Back-translation
- Generating examples with an LLM
- Speech
- Adding noise
- Speed change
Why it matters
Models tend to work well only in situations similar to the data they saw in training. With little data, a model can easily memorize the training data wholesale and then perform worse on new data, which is overfitting. Augmentation shows many variations of the same subject, helping the model focus on 'essential features'. It is especially effective in fields where data is hard to collect and labeling is expensive, such as medical images or photos of defective products.
Things to know
If a transformation changes the correct answer, it does harm. Flipping the digit '6' turns it into '9', and if a negation is dropped while rewording a sentence, the meaning reverses. So which transformations are safe must be judged separately for each dataset and task. Augmentation also only increases variety within the range of the original data; it cannot fill in situations that were missing from the start (for example, no night-time photos at all). Augmented data should not be mixed into evaluation data, so that performance is measured honestly.
Example
When building an AI for defect inspection in a factory, there are far fewer photos of defects than of normal products. Turning the few defect photos into many by rotating, zooming and adjusting brightness, then training on them, produces a model that finds defects well even when lighting or angles differ slightly. For a chatbot intent-classification model, an example like 'I want a refund' may be expanded with an LLM into various phrasings such as 'Can I get my money back?' for training.