“A neural network can compensate for a data augmentation bug, such as forgetting to flip labels when flipping images, by learning to detect the flipped images internally and adjusting its predictions.”