FET: Small: Efficient edge computing using device and network dynamics

The rapid developments of data-intensive applications such as Artificial Intelligence have created significant strains in energy consumption and computing resources. These challenges can be efficiently addressed through computing hardware and algorithm innovations. On the hardware side, the performance of the system is increasingly limited by data movement costs, and new bio-inspired computing architectures can more efficiently process data at significantly lower energy consumption. On the algorithm side, bio-inspired algorithms can process information with much smaller model sizes, leading to further energy and throughput gains. The research will broadly impact neuromorphic computing, AI, and brain-machine interfaces. There are also plans to focus on commercialization of the research results and subsequent tech transfer. Integrating this research with education, the investigator will collaborate with local high schools with large minority student population at annual summer camps hosted by the university, train undergraduate and graduate students, and develop new course modules. This project aims to develop highly efficient bio-inspired edge computing systems that can natively process information at all stacks of the system, based directly on internal device and network dynamics. By leveraging the internal ionic/electronic/thermal dynamic processes in emerging devices, networks and systems can autonomously process spatiotemporal data with high performance, energy efficiency and reliability. Two types of devices, short-term memory memristor and 2nd-order memristor, will be developed and used to directly process temporal inputs and achieve self-learning, respectively. Combined with new network architectures such as reservoir nodes and columnar networks with lateral and feedback connections, the proposed Reservoir Node Networks and Neuromorphic Retina Networks will be able to directly process asynchronous inputs from neuromorphic sensors such as event-based cameras or touch sensors, extract spatiotemporal features at different temporal and spatial scales, and perform object detection and other decisions with unparalleled energy efficiency, latency and robustness. These device and architecture developments will in turn stimulate developments of new bio-inspired algorithms, and enable applications including efficient on-sensor processing, brain-machine interfaces and autonomous systems. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.