Reviews: A Bayesian Data Augmentation Approach for Learning Deep Models

The authors propose a Bayesian approach to handle Data Augmentation (DA) for learning deep models. Data augmentation, in this context, addresses the problem of data sparsity by generating more training samples to make the training process of a (heavily parameterized) Classifier Network (CN) robust. The standard Generative Model (GM) for obtaining the augmented training set is typically trained once and then new data are generated based on this GM and used to estimate the CN. The authors suggest to also learn the GM itself jointly with the CN as new augmented training data are produced - which is straightforward with a Bayesian approach. Using three different data sets and two classifiers, they demonstrate that the Bayesian DA is superior to regular DA, which in turn is superior to not having DA at all (which likely suffers from some overfitting).