Data Augmentation

Applying random transformations to existing training data to create new samples, without collecting anything new.

Data augmentation means expanding a training set by applying random, meaning-preserving transformations to existing samples rather than collecting new ones — a standard way to reduce overfitting and improve robustness. Common image transformations include random cropping, translation, rotation, color jitter, and random occlusion. Because robot data is expensive and scarce, augmentation is especially valuable there, but the transformation must not break the correspondence between the image and the action label — for instance, flipping an image left-right means the left/right direction in the action must be flipped too. RAD (2020) and similar work showed that adding only simple augmentations like random crop and random translation noticeably improves the data efficiency of pixel-based reinforcement learning. More recent work also includes generative data augmentation, which uses generative models to swap backgrounds or objects, and instruction augmentation, which rewrites language instructions.

ExampleRAD added random crop and random translation to the pixel inputs of the DeepMind Control Suite, reaching state-of-the-art data efficiency and final performance at the time, and it also clearly improved test-time generalization on ProcGen.

Also called

Image Augmentation

Related

Overfitting · Generalization · Domain Randomization · Generative Data Augmentation · Instruction Augmentation · Visual Generalization

Sources

Google Machine Learning Glossary: data augmentation
Reinforcement Learning with Augmented Data (RAD, arXiv 2004.14990)