EasyDeepLearn

Why is data augmentation effectively free regularization?

easy

Answer

  • Every augmentation is a prior about invariances the true function satisfies (translations of a cat are still a cat).
  • Applying augmentations enlarges the effective training distribution without labeling cost.
  • Forces the model to be invariant to those transformations, which reduces overfitting to specific instances.
  • Free lunch in vision; needs task-specific care in text and tabular.
Check yourself — multiple choice
  • Augmentation doubles training data at random
  • Encodes invariance priors → larger effective distribution → reduces overfitting for free
  • Same as label smoothing
  • Only for classification

Augmentation = invariance prior → free effective data → regularization.

#augmentation#regularization

Practise Deep Learning

214 interview questions in this topic.

Related questions