Featuring Dr. Eva Dyer, Georgia Institute of Technology

Abstract: Data augmentation is used in almost every facet of machine learning. Despite its widespread use, augmentations are still selected in an ad hoc manner and it's often unclear how to design new augmentations across different tasks and novel domains. In this talk, I will describe a new theoretical framework for characterizing the impact of augmentation on model generalization. Our framework reveals that augmentations induce implicit spectral regularization through a combination of two distinct effects: a) manipulation of the relative proportion of eigenvalues of the data covariance matrix in a training-data-dependent manner, and b) uniformly boosting the entire spectrum of the data covariance matrix through ridge regression. These effects, when applied to popular augmentations, give rise to a wide variety of phenomena, including discrepancies in generalization between over-parameterized and under-parameterized regimes and differences between regression and classification tasks. After describing these theoretical results, we will discuss ways to improve generalization through adaptive augmentations and discuss applications of augmentations in neuroscience.

Bio:  Eva Dyer (she/they) is an Associate Professor in the Coulter Department of Biomedical Engineering at the Georgia Institute of Technology. Dr. Dyer’s research lies at the interface of artificial intelligence and neuroscience, where she aims to both use AI to understand neural computation (AI for Neuro) and use insights from the nervous system to develop robust and lifelong AI systems (Neuro for AI). Eva completed all of her degrees in Electrical & Computer Engineering, obtaining her Ph.D. and M.S. from Rice University and a B.S. from the University of Miami. She is the recipient of a Sloan Fellowship in Neuroscience, NSF CAREER Award, Next Generation Leader Award from the Allen Institute, a McKnight Foundation Technological Innovations in Neuroscience Award, and a CIFAR Azrieli Global Scholar Award.