Abstract
Convolutional neural networks are capable of learning powerfulrepresentational spaces, which are necessary for tackling complex learningtasks. However, due to the model capacity required to capture suchrepresentations, they are often susceptible to overfitting and thereforerequire proper regularization in order to generalize well. In this paper, weshow that the simple regularization technique of randomly masking out squareregions of input during training, which we call cutout, can be used to improvethe robustness and overall performance of convolutional neural networks. Notonly is this method extremely easy to implement, but we also demonstrate thatit can be used in conjunction with existing forms of data augmentation andother regularizers to further improve model performance. We evaluate thismethod by applying it to current state-of-the-art architectures on theCIFAR-10, CIFAR-100, and SVHN datasets, yielding new state-of-the-art resultswith almost no additional computational cost. We also show improved performancein the low-data regime on the STL-10 dataset.