Reviews: Learning to Compose Domain-Specific Transformations for Data Augmentation

This paper addresses an interesting and new problem to augment training data in a learnable and principled manner. Modern machine learning systems are known for their'hunger for data' and until now state-of-the-art approaches have relied mainly on heuristics to augment labeled training data. This paper tries to reduce the tedious task of finding a good combination of data augmentation strategies with best parameters by learning a sequence of best data augmentation strategies in a generative adversarial framework while working with unsupervised data. The motivation behind the problem is to reduce human labor without compromising the final discriminative classification performance. The problem formulation is pretty clear from the text.