SHF: Small: Efficient and Accurate Learning with Low-Precision Components: A Cortex-Inspired Approach

Arizona State University, Scottsdale AZ

Investigators

Abstract

Achieving high performance and high-energy efficiency with a small footprint is a central challenge of computer engineering. This project targets to develop technologies with cutting-edge nanoscale devices towards a self-learning chip. It will be integrated with front-end sensors, process the information in real-time, and consume ultra-low energy. The success is likely to have an impact on the society, bringing broad benefits to multiple emerging applications, mobile vision and autonomous vehicles to name a few. The interdisciplinary nature of this project, as well as the frequent interaction with industry, will provide an ideal platform for education and training of state-of-the-art science and technology. It will improve the knowledge base of intelligent system design through new curriculum development, engaging undergraduate and minority students in research and practice, and participating in outreach programs that are customized for K-12 students. Furthermore, this project will advocate the web-based interface and workshops to disseminate the latest research outcome. Microprocessors have been a ubiquitous and vitally important part in our modern-day life. However, they are facing severe issues in artificial intelligent systems, which require tremendous amount of energy and data to train and operate the sophisticated algorithm. On the contrary, animal brains at various sizes achieve remarkable feats of learning and accuracy at energy costs much lower than human-engineered systems. Therefore, the central theme of this project is to transfer the latest knowledge of the structure and function of brains into neuromorphic design, generate novel insights for improvement of the engineered system, and achieve high accuracy and high energy efficiency despite the severe precision constraints of the nanoscale components. These neurobiological principles include approximate learning rules with low-precision synapses, neural motifs of excitation and inhibition, and hierarchical network models. The goal is to accomplish complex computation with much less data volume and resources, and promise magnitudes of improvement in energy efficiency and performance than microprocessors today.