RandAugment is an automated data augmentation technique designed to improve the performance of machine learning models, particularly in computer vision tasks. It simplifies the process of applying random transformations to training images, helping models generalize better by exposing them to a wider variety of data variations. Unlike older methods that require complex tuning of augmentation policies, RandAugment reduces the search space to two hyperparameters, making it easier to implement and scale across datasets.
The method works by randomly selecting a fixed number of transformations (denoted as N) from a predefined list of image-processing operations, such as rotation, shearing, or color adjustments. Each transformation is applied with a uniform intensity (M), which controls the magnitude of the effect (e.g., how much to rotate an image). For example, if N=2 and M=9, RandAugment might apply a 30-degree rotation (scaled by M) followed by a color inversion. The key innovation is that both the selection of transformations and their order are randomized per image, ensuring diversity without requiring handcrafted policies. Developers typically define a list of 10–20 base operations (e.g., flipping, adjusting brightness, adding noise), and the system handles the rest stochastically during training.
RandAugment’s strength lies in its simplicity and efficiency. Earlier approaches like AutoAugment used reinforcement learning to find optimal policies, which was computationally expensive. RandAugment eliminates this overhead by relying on randomness and shared magnitude values, while still achieving comparable or better results. For instance, in a practical implementation, a developer might set N=3 and M=12 for a dataset, leading to combinations like shear + solarize + contrast adjustments. This approach reduces the risk of overfitting to a fixed augmentation sequence and works well across tasks without dataset-specific tuning. By focusing on minimal hyperparameters and leveraging randomness, RandAugment provides a flexible, low-maintenance solution for enhancing model robustness.