Robust Machine Learning Methods and Applications

Abstract

This dissertation introduces a variety of methods motivated from the use of robust losses in machine learning and deep neural networks and explores their applications. Topics include novel linear and nonlinear approaches for imbalanced data classification, their applications in functional magnetic resonance imaging, robust deep neural networks and the application in adversarial attacks. Robustness is mainly used to test whether the model can still maintain accuracy of judgment in the face of small changes in input data, that is, whether the performance of the model is stable in the face of certain changes. The level of robustness directly impacts the generalization ability of machine learning models. In this dissertation, we have used different robust losses to propose, analyze, and evaluate models. First, robust support vector classifier loss was adapted to propose two new classifiers, the pairwise robust support vector machine and its kernel counterpart, which are shown more efficient for imbalanced data classification. Then, correntropy loss was used to propose two robust deep neural networks and their two-stage implementations. Finally, correntropy loss was introduced for adversarial attack to further analyze the robustness of an image classification model. Simulations and case studied show that with the introduction of robust losses into machine learning and deep learning, the newly proposed models have better performance than traditional models and can improve model robustness.