How Much Data Are Augmentations Worth? An Investigation into Scaling Laws, Invariance, and Implicit Regularization
This work addresses the problem of understanding and optimizing data augmentations for machine learning practitioners, offering insights into their mechanisms and value, though it is incremental in building on existing knowledge.
The paper investigates why data augmentations are effective, finding that in out-of-distribution scenarios, diverse but inconsistent augmentations can be more valuable than additional real data, and that augmentations encouraging invariances are particularly beneficial on small to medium training sets by flattening the loss landscape.
Despite the clear performance benefits of data augmentations, little is known about why they are so effective. In this paper, we disentangle several key mechanisms through which data augmentations operate. Establishing an exchange rate between augmented and additional real data, we find that in out-of-distribution testing scenarios, augmentations which yield samples that are diverse, but inconsistent with the data distribution can be even more valuable than additional training data. Moreover, we find that data augmentations which encourage invariances can be more valuable than invariance alone, especially on small and medium sized training sets. Following this observation, we show that augmentations induce additional stochasticity during training, effectively flattening the loss landscape.