Hardware-Software Co-design of Sustainable AI: Neuromorphic Computing and Beyond

PASSWORD: 387941

  The relentless pursuit for AI is rapidly producing advanced algorithms and models, and abundant data, which on the other hand causes unsustainable power consumption projections. For sustainability, there is an urgent demand for better power-efficiency in both hardware and algorithms. In this dissertation, I discuss novel hardware-aware algorithms and algorithm-aware hardware designs that maximize each other’s utility. The approach covers the full stack of AI computing, including: (1) devices that natively possess information through internal dynamics, (2) device physics-enabled algorithms and architectures for brain-inspired neuromorphic systems and AI accelerators, and (3) cognitive algorithms enabling more efficient and better AI inference.

   The target of my research is benefitting from both dynamic devices and brain-inspired and cognitive algorithms for AI computing systems to achieve both high performance and high energy-efficiency. Smartly leveraging device dynamics will eliminate unnecessity arising from digital-processing of analog and non-linear functions that are natural to human but expensive to digital computers. Inspiration from the brain grants biological learning abilities to build intelligent systems, without astronomical training costs. Cognitive algorithms allow AI models to adapt to what they encounter. These innovations collaboratively help accomplish “Artificial Intelligence” as the name truly means, beyond the current stage of “machine intelligence”.

CHAIR: Professor Wei D. Lu