Data augmentation has played a central role in many of the recent advances in supervised, self-supervised, and semi-supervised learning. Test-time corruptions, an analog of data augmentation at test time, were instrumental in recent attempts at characterizing the robustness and out-of-domain generalization ability of deep learning models. I will present a summary of recent discoveries about the generalization ability of vision models using data augmentation and test-time corruptions. I will also include practical recommendations on how data augmentation can be utilized to produce more accurate and more robust vision models.
Ekin Dogus Cubuk is a Senior Research Scientist at Google Brain where he works on deep learning and its applications to physical sciences. He received his Ph.D. from Harvard University where he studied disordered solids using machine learning. He has recently focused on using data augmentation, self-training, and test-time corruptions to study the generalization ability of deep learning models.