Data augmentation is the practice of producing new data for machine learning by processing existing data. For example, if the goal is to train a cat-vs-dog image classifier, then whether or not the image is reflected left or right does not matter, which means we get two times more data.