- you can manipulate original training data for new altered synthetic examples - i.e flip image, distort - not as good as having an actual new piece of training data though - you do this when you need more data - helps prevent [[Bias and Variance|Overfitting]], teaches model that distorted forms are same object