Augmax

AugMax is a data augmentation strategy that adversarially combines multiple augmentation chains to create the most challenging augmented sample — finding the worst-case mixture of augmentations that maximally increases the training loss, providing robustness training.

How Does AugMax Work?

• Multiple Chains: Apply $K$ different augmentation chains to the same input (e.g., $K = 3$).
• Adversarial Mixture: Find the convex combination $\sum_k w_k \cdot \text{Aug}_k(x)$ that maximizes the loss.
• Train: Train the model on this worst-case augmented sample.
• Paper: Wang et al. (2021).

Why It Matters

• Adversarial Augmentation: Goes beyond random augmentation by actively finding the hardest combination.
• Robustness: Improves both clean accuracy and corruption robustness (ImageNet-C, ImageNet-P).
• Principled: The adversarial mixture is a principled way to explore the augmentation space efficiently.

AugMax is augmentation as an adversary — finding the hardest possible augmentation mixture to create maximally challenging training samples.