Data Augmentation

Augmentation alters existing examples through operations such as cropping, rotation or controlled perturbation, rather than collecting independent observations. The transformation encodes an assumption about invariance: the target should remain valid despite the change. For detection or segmentation, spatial labels must be transformed consistently with the image. Some changes require updating the label, and others destroy essential information. This differs from general synthetic data generation, which can create entirely new examples, although the boundary can overlap. Augmentation changes the training distribution and should reflect meaningful deployment variation rather than simply maximizing the number of apparently different samples.

The practitioner identifies expected nuisance variation and selects transformations with justified ranges. They inspect transformed examples and targets, then evaluate the policy on held-out data that has not been artificially made easier. Useful artifacts include a versioned augmentation configuration and tests for label alignment. Random transformations occur within training boundaries, while validation transformations follow a defined protocol. The team compares task performance across relevant conditions and checks whether the policy damages rare cases or teaches invariances that the application should not have.

More augmented samples are not more independent evidence. Invalid transformations can corrupt labels, and a model may learn artifacts introduced by the augmentation process. Aggressive policies can erase details needed for the task. The quality test checks both transformation validity and held-out behavior under realistic conditions. Augmentation cannot replace missing population coverage or justify evaluating on near-duplicates of training examples generated from the same source.