[Paper Review] RandAugment: Practical data augmentation with no separate search.

Ekin D. Cubuk, Barret Zoph+2 more|arXiv (Cornell University)|Sep 30, 2019

Advanced Neural Network ApplicationsComputer Science51 references183 citations

TL;DR

RandAugment proposes a practical, search-free data augmentation method that directly optimizes augmentation policies on the target model and dataset, eliminating the need for proxy tasks. It achieves state-of-the-art performance on CIFAR-10, CIFAR-100, SVHN, and ImageNet (85.0% top-1 accuracy), surpassing prior learned methods like AutoAugment while enabling dynamic regularization strength adjustment based on model and dataset size.

ABSTRACT

Recent work has shown that data augmentation has the potential to significantly improve the generalization of deep learning models. Recently, learned augmentation strategies have led to state-of-the-art results in image classification and object detection. While these strategies were optimized for improving validation accuracy, they also led to state-of-the-art results in semi-supervised learning and improved robustness to common corruptions of images. One obstacle to a large-scale adoption of these methods is a separate search phase which significantly increases the training complexity and may substantially increase the computational cost. Additionally, due to the separate search phase, these learned augmentation approaches are unable to adjust the regularization strength based on model or dataset size. Learned augmentation policies are often found by training small models on small datasets and subsequently applied to train larger models. In this work, we remove both of these obstacles. RandAugment may be trained on the model and dataset of interest with no need for a separate proxy task. Furthermore, due to the parameterization, the regularization strength may be tailored to different model and dataset sizes. RandAugment can be used uniformly across different tasks and datasets and works out of the box, matching or surpassing all previous learned augmentation approaches on CIFAR-10, CIFAR-100, SVHN, and ImageNet. On the ImageNet dataset we achieve 85.0% accuracy, a 0.6% increase over the previous state-of-the-art and 1.0% increase over baseline augmentation. On object detection, RandAugment leads to 1.0-1.3% improvement over baseline augmentation, and is within 0.3% mAP of AutoAugment on COCO. Finally, due to its interpretable hyperparameter, RandAugment may be used to investigate the role of data augmentation with varying model and dataset size.

Motivation & Objective

• To eliminate the need for a separate proxy search phase in learned data augmentation, which increases training complexity and computational cost.
• To enable automatic adaptation of regularization strength based on model and dataset size, overcoming the limitation of fixed policies trained on small proxies.
• To develop a unified, plug-and-play augmentation strategy that works effectively across diverse tasks and datasets without retraining or reconfiguration.
• To improve generalization and robustness in image classification and object detection while maintaining interpretability through a single, intuitive hyperparameter.
• To enable systematic study of data augmentation's impact by varying regularization strength in relation to model and dataset scale.

Proposed method

• RandAugment uses a simple, fixed parameterization: it applies a random subset of K out of N possible augmentations to each input image, with K being a tunable hyperparameter.
• The augmentation operations are selected uniformly at random from a predefined search space, removing the need for reinforcement learning or neural architecture search.
• The policy is trained directly on the target model and dataset, using standard training loops without proxy tasks or separate search stages.
• The strength of regularization is controlled by the hyperparameter K, which determines the number of augmentations applied per image and can be tuned per model or dataset size.
• The method leverages the invariance of random augmentation selection to enable consistent performance across different training regimes.
• It uses a differentiable search space where operations are applied with equal probability, allowing gradient-free optimization to be avoided.

Experimental results

Research questions

• RQ1 Can data augmentation policies be learned effectively without a separate search phase, directly on the target model and dataset?
• RQ2 Can a simple, random selection of augmentations achieve state-of-the-art performance compared to learned policies that require complex search?
• RQ3 How does the regularization strength of data augmentation scale with model and dataset size, and can it be dynamically adjusted?
• RQ4 Can a single, interpretable hyperparameter (K) effectively control augmentation strength across diverse tasks and datasets?
• RQ5 Does a search-free, random augmentation strategy outperform or match the performance of reinforcement learning-based methods like AutoAugment on standard benchmarks?

Key findings

• RandAugment achieves 85.0% top-1 accuracy on ImageNet, representing a 0.6% improvement over the previous state-of-the-art and a 1.0% improvement over standard baseline augmentation.
• On CIFAR-10, CIFAR-100, and SVHN, RandAugment matches or exceeds the performance of all prior learned augmentation methods.
• In object detection on COCO, RandAugment improves mAP by 1.0–1.3% over baseline augmentation and is within 0.3% mAP of AutoAugment.
• The method eliminates the need for a separate search phase, reducing training complexity and enabling direct training on the target model and dataset.
• The hyperparameter K provides interpretable control over regularization strength, allowing systematic study of augmentation effects across model and dataset sizes.
• RandAugment is effective out-of-the-box across diverse tasks and datasets, requiring no reconfiguration or retraining for different settings.