[Paper Review] AutoAugment: Learning Augmentation Policies from Data

Ekin D. Cubuk, Barret Zoph+3 more|arXiv (Cornell University)|May 24, 2018

Advanced Neural Network Applications65 references732 citations

TL;DR

AutoAugment automatically searches for effective data augmentation policies using a controller RNN and reinforcement learning, achieving state-of-the-art results on multiple datasets and transferable policies across datasets.

ABSTRACT

Data augmentation is an effective technique for improving the accuracy of modern image classifiers. However, current data augmentation implementations are manually designed. In this paper, we describe a simple procedure called AutoAugment to automatically search for improved data augmentation policies. In our implementation, we have designed a search space where a policy consists of many sub-policies, one of which is randomly chosen for each image in each mini-batch. A sub-policy consists of two operations, each operation being an image processing function such as translation, rotation, or shearing, and the probabilities and magnitudes with which the functions are applied. We use a search algorithm to find the best policy such that the neural network yields the highest validation accuracy on a target dataset. Our method achieves state-of-the-art accuracy on CIFAR-10, CIFAR-100, SVHN, and ImageNet (without additional data). On ImageNet, we attain a Top-1 accuracy of 83.5% which is 0.4% better than the previous record of 83.1%. On CIFAR-10, we achieve an error rate of 1.5%, which is 0.6% better than the previous state-of-the-art. Augmentation policies we find are transferable between datasets. The policy learned on ImageNet transfers well to achieve significant improvements on other datasets, such as Oxford Flowers, Caltech-101, Oxford-IIT Pets, FGVC Aircraft, and Stanford Cars.

Motivation & Objective

• Motivate automatic discovery of augmentation strategies to improve image classifier generalization.
• Develop a search space for policies consisting of sub-policies with sequential image operations.
• Use a controller (RNN) and policy-gradient reinforcement learning to optimize validation accuracy.
• Demonstrate direct application (AutoAugment-direct) and cross-dataset transfer (AutoAugment-transfer) of learned policies.

Proposed method

• Define a policy as 5 sub-policies, each with two image operations in sequence.
• Each operation has a probability of application and a magnitude value.
• Use a controller RNN to sample policies and train a child model to measure validation accuracy as reward.
• Optimize the controller with Proximal Policy Optimization (PPO) using the policy gradient from rewards.
• Concatenate the best 5 policies into a final 25-sub-policy policy for training datasets.

Experimental results

Research questions

• RQ1 Can automatically learned augmentation policies improve image classifier performance beyond manually designed augmentations?
• RQ2 Are policies found on one dataset transferable to other datasets and architectures?
• RQ3 What is the impact of policy diversity and stochastic application on generalization?
• RQ4 How does AutoAugment compare to prior automated augmentation methods in accuracy and transferability?

Key findings

• AutoAugment achieves state-of-the-art error rates on CIFAR-10 (1.5%), CIFAR-100 (10.7%), and SVHN (1.0%), and Top-1 accuracy 83.5% on ImageNet.
• Policies learned on ImageNet transfer to FGVC datasets and improve generalization without additional data.
• On CIFAR-10, AutoAugment yields substantial gains over prior methods across multiple architectures.
• The transfered ImageNet policy improves Stanford Cars and FGVC Aircraft by 1.2% and 1.8% respectively.
• Policies discovered directly on reduced CIFAR-10 also improve full CIFAR-10/100 and SVHN results, showing robustness across scales.