Data Augmentation

Data augmentation artificially expands a training set by applying label-preserving transformations to existing examples — random crops, flips, rotations, or color jitter for images. A photo of a cat flipped horizontally is still obviously a cat, so this manufactures new, valid training examples at no extra labeling cost, and is standard practice in nearly every CNN training pipeline as a cheap defense against overfitting.

How it works

Augmentation runs on the fly in the input pipeline, so each epoch sees a different random perturbation of the same image and the model never fits a fixed arrangement of pixels. A typical image policy composes RandomResizedCrop, RandomHorizontalFlip, and color jitter, then normalization. Two properties matter: the transform must preserve the label, and the randomness must be resampled per example rather than fixed once per epoch. Stronger policies mix examples instead of merely perturbing them — mixup blends two images and their labels linearly, CutMix pastes a patch of one image into another and mixes the labels in proportion to area. Functionally this is a regularizer, and it belongs to the training split only.

When it breaks

• Applying random transforms to validation or test data silently corrupts your evaluation split; only deterministic resize and normalize belong there.
• An overly aggressive policy destroys signal — a hard crop can remove the object entirely, leaving a confidently mislabeled example.
• The transforms encode assumptions. Horizontal flips are fine for natural photos and wrong for medical scans, where laterality is diagnostic.
• CPU-side augmentation routinely becomes the training bottleneck, leaving the accelerator idle between batches.

See also: Overfitting, CNN

Learn more: Computer Vision