Meta Generative Image and Text Data Augmentation Optimization

Abstract This paper proposes a method called Meta Generative Data Augmentation Optimization (MGDAO) to overcome limited types of operations for the policy-based automatic data augmentation method. While traditional data augmentation methods have relied on expert intuition to determine effective transformations, recent approaches have attempted to generate data augmentation strategies automatically. However, these automatic methods can still suffer from limited operation sets and difficulty training conditional generative models. To address these issues, MGDAO replaces the limited operations space in the AutoAugment series with a deep-style generator and replaces the discriminator in a generative adversarial model with the validation loss of the target model. These replacements released fixed image operations and made MGDAO useful for sequential data. The generator learns to transform the data from the training domain to the validation data domain. It is further used to generate augmented samples to train the target model and reduce the validation loss. Experiments on classification benchmarks of few-shot image and text-based datasets show that MGDAO achieves competitive results compared to data auto-augmentation methods.