Improved Regularization of Convolutional Neural Networks with Cutout

Convolutional neural networks can learn the powerful representational spaces needed for complex tasks, but the large model capacity this requires makes them prone to overfitting, so they need effective regularization to generalize well. This paper proposes cutout, a simple regularization technique that randomly masks out square regions of the input image during training. The method is extremely easy to implement and can be combined with existing forms of data augmentation and other regularizers to further improve performance.

Applied to then-current state-of-the-art architectures, cutout improves robustness and overall accuracy, producing new state-of-the-art test errors of 2.56% on CIFAR-10, 15.20% on CIFAR-100, and 1.30% on SVHN. Its simplicity, low cost, and compatibility with other techniques make it an easy addition to image-classification training pipelines.