Understanding Data Manipulation and How to Leverage it To Improve Generalization

Abstract

Augmentations and other transformations of data, either in the input or latent space, are a critical component of modern machine learning systems. While these techniques are widely used in practice and known to provide improved generalization in many cases, it is still unclear how data manipulation impacts learning and generalization. To take a step toward addressing the problem, this thesis focuses on understanding and leveraging data augmentation and alignment for improving machine learning performance and transfer. In the first part of the thesis, we establish a novel theoretical framework to understand how data augmentation (DA) impacts learning in linear regression and classification tasks. The results demonstrate how the augmented transformed data spectrum plays a key role in characterizing the behavior of different augmentation strategies, especially in the overparameterized regime. The tools developed in this aim provide simple guidelines to build new augmentation strategies and a simple framework for comparing the generalization of different types of DA. In the second part of the thesis, we demonstrate how latent data alignment can be used to tackle the domain transfer problem, where training and testing datasets vary in distribution. Our algorithm builds upon joint clustering and data-matching through optimal transport, and outperforms the pure matching algorithm baselines in both synthetic and real datasets. Extension of the generalization analysis and algorithm design for data augmentation and alignment for nonlinear models such as artificial neural networks and random feature models are discussed. This thesis provides tools and analyses for better data manipulation design, which benefit both supervised and unsupervised learning schemes.