autoaugment

**AutoAugment** is a **learned data augmentation strategy that uses reinforcement learning to search for the best augmentation policy** — discovering which combinations and magnitudes of image transformations maximize validation accuracy for a given dataset. **How Does AutoAugment Work?** - **Search Space**: Each policy = 5 sub-policies. Each sub-policy = 2 transformations, each with probability and magnitude. - **Controller**: An RNN controller proposes augmentation policies. - **Reward**: The policy is evaluated by training a small child model — validation accuracy is the reward. - **Transfer**: Policies found on ImageNet transfer well to other datasets. - **Paper**: Cubuk et al. (2019, Google Brain). **Why It Matters** - **Learned Augmentation**: Demonstrated that augmentation strategies can be learned, not just hand-designed. - **Accuracy Boost**: +0.4-1.0% on ImageNet, larger gains on smaller datasets (CIFAR-10, SVHN). - **Expensive**: The search process requires thousands of GPU hours — motivating RandAugment. **AutoAugment** is **NAS for data augmentation** — using reinforcement learning to discover the optimal augmentation recipe for any dataset.