Self-supervised regression learning (SSRL)
Regression that predicts real/complex-valued output from real/complex-valued input, is a central part of applications using computational imaging and computer vision technologies. Yet, self-supervised learning in the regression setup, i.e., SSRL, has not been studied/understood. This project studies SSRL incorporating domain knowledge, for a range of applications in imaging, image processing, and computer vision. The project is being supported by HCF (project end date: 6/1/22).
Fast iterative neural network (INN)
Conventional regression AI systems in imaging, image processing, and computer vision can have over-fitting risks. INNs such as Momentum-Net and BCD-Net resolve the limitation by using domain knowledge such as imaging physics and image formation models in an iterative fashion. This project develops faster INNs, to enable near real-time image reconstruction. The project is being supported by the PI's seed funds (project end date: 12/31/22).
End-to-end (E2E) autonomous driving
E2E driving is an emerging regression AI technology that directly maps input raw data such as image(s) – one end – to vehicle control signals – the other end. This project investigates E2E driving AI systems with multi-modal imagers to develop self-driving cars optimized for Hawaii. The project is in collaboration with Dr. David Ma (Interim Associate Dean at UHM CoE and Professor of CEE at UHM) and being supported by HDOT (project end date: 7/31/22).
There exist several practical changes in high-speed autonomous driving, including vibrations from a vehicle and motion blur in images/data captured by cameras, Lidar, etc. This project will develop and implement AI systems for high-speed self-driving cars. This is in collaboration with UH's AI Racing Tech Team (project end date: 7/31/22).