Public Lecture #2

"Domain-Specific Acceleration Targeting AI for Science and AI for Medicine"

Prof. Wayne LUK (Imperial College London)

Abstract

Special-purpose computer architectures have shown significant promise in recent years. Such architectures can benefit from domain-specific customization to enhance their capabilities, resulting in superior performance for demanding applications. This talk describes advances in domain-specific acceleration for two important applications involving artificial intelligence: AI for science and AI for medicine.

To illustrate accelerator development targeting AI for science, I will describe our collaboration with high-energy physics researchers at CERN in developing low-latency deep learning accelerator architectures. In particular, I will present a Graph Neural Network accelerator capable of achieving sub-microsecond latency to support online event selection in the Level-1 triggers at the CERN Large Hadron Collider experiments. Our efforts to automate the development of such accelerators based on meta-programming techniques will also be covered.

To illustrate accelerator development targeting AI for medicine, I will describe our collaboration with adaptive radiotherapy researchers at the Institute of Cancer Research in developing accelerator architectures with improved trustworthiness for healthcare applications. In particular, I will present a Bayesian Neural Network accelerator capable of efficient uncertainty estimation for Magnetic Resonance Imaging analysis. Our efforts to automate the development of such accelerators based on a novel algorithm-hardware co-optimization flow will also be covered.