Data Augmentation
Creating extra training examples by changing existing ones in harmless ways, such as flipping, cropping or recoloring images.
Also augmentation · random crops · flips
A cat photo flipped left to right is still a cat. Showing the model both versions teaches it that direction does not matter, without anyone taking a new photo.
Where it fits
+All relationships as a list
How it works
Neural networks need lots of examples, and labeled examples are expensive. Augmentation stretches a dataset by applying transformations that change the pixels but not the meaning:
- Geometric: flips, small rotations, random crops and zooms.
- Color: brightness, contrast, saturation shifts.
- Noise and blur: to mimic cheap cameras and motion.
- Mixing: blend two images and their labels (mixup), or paste a patch of one onto another (CutMix).
Each training epoch sees a slightly different version of every image, so the network cannot simply memorize pixels. That makes augmentation a form of regularization against overfitting. AlexNet relied on random crops and flips to train on ImageNet.
The rule: only use transformations that keep the label true for your task.
Everyday use cases
- Case 01
Small datasets
A clinic with only a few hundred labeled scans can multiply them with rotations and brightness changes.
- Case 02
Robust phone cameras
Training with blur, noise and different lighting helps models work on real snapshots.
Famous examples
Strengths, weaknesses and when to use it
Strengths
- Free extra data, often a large accuracy gain.
- Teaches invariances you know should hold.
Weaknesses
- Transformations must not change the label (a flipped "6" becomes a "9"-like shape; flipped text is no longer text).
When to use itAlmost always for image models, and in adapted forms for audio and text.
Check yourself
- Q01
Which augmentation would be a bad idea for reading digits?
Show answer
C: Flipping upside down An upside-down 6 looks like a 9, so the label would no longer be correct.