Transformations such as cropping, noise, paraphrasing, or simulation expose the model to plausible variation. The transformation must match the domain, because unrealistic changes can teach incorrect invariances.
Data augmentation creates modified training examples that preserve the intended label or meaning.
Transformations such as cropping, noise, paraphrasing, or simulation expose the model to plausible variation. The transformation must match the domain, because unrealistic changes can teach incorrect invariances.