google-research/augmix

What it solves

AugMix is designed to improve the robustness and uncertainty calibration of image classification models. It addresses the problem of models failing when encountering images that have been corrupted or shifted in distribution from the training data (e.g., ImageNet-C).

How it works

The technique mixes multiple augmented versions of an image and enforces consistent embeddings of these augmented images. It uses a Jensen-Shannon Divergence (JSD) consistency loss to ensure the model's predictions remain consistent between two different augmentations of the input image and the clean image itself.

Who it’s for

Machine learning engineers and researchers focusing on computer vision and image classification who want to increase their model's resilience to image corruptions without needing extensive hyperparameter tuning for the augmentation process.

Highlights

• Plug-and-play: Does not require the manual tuning typically associated with methods like random cropping or CutOut.
• State-of-the-art performance: Achieves high robustness and uncertainty estimation on benchmarks like ImageNet-C and ImageNet-P.
• Reference implementation: Provides a NumPy implementation of the augmentation method and PyTorch re-implementations for CIFAR and ImageNet datasets.
• Consistency Loss: Employs JSD loss to stabilize predictions across different augmented versions of the same image.