Adaptive AutoAugment

Abstract

This study deals with the algorithm that searches optimal data augmentation policies which can improve the performance of the machine learning algorithm in vision related tasks. Data augmentation solves the data shortage in machine learning algorithm and induces robust training in image process problems resulting in the improved generalization of target tasks. Recently, human-designed data augmentation has been replaced by automatically the searched augmentation policy in the pre-defined search space of data augmentation. AutoAugment can significantly improve the accuracy of the image classification tasks by searching the optimal augmentation policy for the dataset with reinforcement learning (RL). In previous studies, however, the same augmentation policy is applied throughout the entire dataset so that every data in the dataset is pre-processed with the same augmentation policy. In this study, the augmentation policy search method for individual data, Adaptive AutoAugment, has been proposed. Adaptive AutoAugment has applied different augmentation policy throughout learning progress even with the same data, while consuming the least computational resources. It has improved the performance of image classification on CIFAR-10/100 with the 2.5 times fewer computational resources than the most efficient method.