How does data augmentation improve the performance and generalization of machine learning models, particularly in image and text classification tasks?