Data Augmentation

Data augmentation creates changed versions of training examples that should preserve the label or target.

For images, that might mean small crops, flips, or color changes. For other data, the safe transformations depend on the domain.

The principle is not "change data randomly." The principle is "show the model variations that should not change the answer."

This chapter only introduces augmentation as a generalization idea. Full augmentation strategy belongs with the domain. Image augmentation belongs with computer vision. Text augmentation is more delicate because small wording changes can alter meaning.

Used well, augmentation can reduce overfitting by making the training set express more of the variation expected at test time.

Exercise: Safe augmentation idea

Enter 1 if augmentation should preserve the target, or 2 if augmentation should intentionally change the target.

Your answer

Compute it first, then check your number.

Exercise: Dataset size effect

If augmentation turns each of 100 examples into 3 variants, how many training variants are available?

Your answer

Compute it first, then check your number.