Data augmentation is a technique used in machine learning to increase the size of the dataset by perturbing or augmenting the existing data to create variations that can improve the performance of models. By using this technique, the neural network can output a vector of features that represent different crops of the image with different lighting, contrast or colors.