(Georgia Institute of Technology, 2020-07-27)
Liu, Weiyang
How to efficiently learn discriminative deep features is arguably one of the core problems in deep learning, since it can benefit a lot of downstream tasks such as visual recognition, object detection, semantic segmentation, etc. In this dissertation, we present a unified deep representation learning framework on hypersphere, which inherently introduces a novel hyperspherical inductive bias into deep neural networks. We show that our framework is well motivated from both empirical observations and theories. We discuss our framework from four distinct perspectives: (1) learning objectives on hypersphere; (2) neural architectures on hypersphere; (3) regularizations on hypersphere; (4) hyperspherical training paradigm. From the first three perspectives, we explain how we can utilize the idea of hyperspherical learning to revisit and reinvent corresponding components in deep learning. From the last perspective, we propose a general neural training framework that is heavily inspired by hyperspherical learning. We conduct comprehensive experiments on many applications to demonstrate that our deep hyperspherical learning framework yields better generalization, faster convergence and stronger adversarial robustness compared to the standard deep learning framework.