RandAugment

Unlike prior methods like AutoAugment, which use reinforcement learning or neural architecture search to find an optimal policy, RandAugment completely removes the separate, computationally expensive search phase. The policy is defined by just two hyperparameters: N (the number of transformations to apply sequentially) and M (a shared magnitude for all transformations). This makes it vastly more efficient to deploy.

For each training sample, RandAugment constructs an augmentation chain by:

• Randomly selecting N transformations** from a predefined set (e.g., identity, shear, translate, rotate, autoContrast, invert, equalize, solarize, posterize, contrast, color, brightness, sharpness, cutout).
• Applying each selected transformation in sequence.
• Sampling the intensity magnitude for each operation uniformly from a range defined by M. This uniform randomness provides a simple yet highly effective form of regularization.

The entire policy's complexity is controlled by just two intuitive parameters, making hyperparameter tuning straightforward:

• N (Number of Transformations): Typically between 1 and 3. Controls the depth of the augmentation chain.
• M (Magnitude): An integer (e.g., 0 to 30) that maps to a severity range for all operations. A higher M applies more aggressive transformations. This simplicity allows RandAugment to scale effectively to large datasets and models where per-dataset search is prohibitive.

By removing the search phase, RandAugment reduces the computational overhead of automated augmentation from thousands of GPU hours to near-zero. The policy:

• Does not require a proxy task or a held-out validation set for policy optimization.
• Adds minimal overhead during training, as random selection and application of transformations are computationally cheap.
• Enables consistent performance across diverse tasks (e.g., ImageNet, COCO, SVHN) with the same small set of hyperparameters.

The random application of diverse transformations acts as a powerful regularizer, forcing the model to learn more robust and invariant features. Key benefits include:

• Reduced overfitting on small and medium-sized datasets.
• Enhanced performance on out-of-distribution and corrupted data benchmarks.
• Consistent accuracy gains across computer vision tasks like image classification, object detection, and semantic segmentation when compared to baseline augmentations.

While originally designed for images, RandAugment's principles are foundational for Multimodal Data Augmentation (MMDA). Its core ideas inform strategies for augmenting paired data:

• Synchronized Augmentation: Applying geometrically consistent transformations (e.g., the same crop) to aligned image-text or image-audio pairs.
• Modality-Agnostic Policy: The concept of a simple, random policy can be extended to other modalities (e.g., applying speed perturbation or time masking to audio, or synonym replacement to text) within a coordinated framework.

The overarching field where algorithms automatically discover optimal augmentation policies. Unlike manual tuning, these methods use search strategies like reinforcement learning or population-based training to find the best sequence and magnitude of transformations for a given dataset and task. RandAugment simplifies this by removing the search phase, using a fixed policy with random selection.

A predefined set of rules governing how training data is transformed. A policy specifies:

• The library of operations (e.g., Rotate, Shear, ColorJitter).
• The probability of applying each operation.
• The magnitude range for each operation.

RandAugment's policy is defined by two hyperparameters: N (number of operations to apply) and M (a shared magnitude for all operations), which are uniformly sampled.

The direct predecessor to RandAugment. AutoAugment uses reinforcement learning to search over a discrete space of augmentation policies, creating a dataset-specific policy. This search is computationally expensive. RandAugment was proposed as a simplified, search-free alternative, demonstrating that random selection with uniform magnitude sampling often matches or exceeds the performance of learned policies.

An inference-time technique that applies multiple random augmentations (like those in RandAugment) to a single test sample. The model makes a prediction for each augmented version, and the results are aggregated (e.g., averaged) to produce a final, more robust prediction. This reduces variance and can improve accuracy, especially on corrupted or out-of-distribution data.

Advanced, label-mixing augmentation techniques often used alongside or compared to RandAugment.

• Mixup: Creates virtual samples via linear interpolation between two images and their labels.
• CutMix: Cuts and pastes a patch from one image onto another, mixing the labels proportionally.

These methods encourage smoother decision boundaries. RandAugment is frequently combined with them in a training pipeline for compounded robustness gains.

A multimodal-specific technique crucial for data with aligned modalities (e.g., video+audio, image+caption). It ensures that the same geometric transformation (like a crop or rotation) is applied identically to all modalities in a paired sample. This preserves cross-modal alignment. While RandAugment is modality-agnostic, applying it in a multimodal context requires synchronized execution to maintain semantic relationships.