Succinct and Assured Machine Learning: Training and Execution

One of the key challenges facing wide-scale adoption of Machine Learning (ML) is making the existing models more scalable, resource efficient, and safe. This challenge is especially apparent on embedded edge devices where memory storage, battery life, and communication bandwidth are limited. A recent popular line of research has focused on performance optimization and ML acceleration using hardware and software co-design to address the scalability and resource efficiency issues. Safety consideration is usually an afterthought if considered at all. My dissertation advances the state-of-the-art in this growing field by advocating a holistic co-design approach which not only includes hardware and software but also considers the geometry of the data, the learning model, as well as safety concerns (e.g., robustness against adversarial attacks and intellectual property infringement).