Improved Mixed-Example Data Augmentation
share
In order to reduce overfitting, neural networks are typically trained with data augmentation, the practice of artificially generating additional training data via label-preserving transformations of existing training examples. Recent work has demonstrated a surprisingly effective type of non-label-preserving data augmentation, in which pairs of training examples are averaged together. In this work, we generalize this "mixed-example data augmentation", which allows us to find methods that improve upon previous work. This generalization also reveals that linearity is not necessary as an inductive bias in order for mixed-example data augmentation to be effective, providing evidence against the primary theoretical hypothesis from prior work.
READ FULL TEXT