New Physics Model Guided Machine Learning:  A physics model is a validated low-dimensional model of causality with critical properties such as provable stability in control applications. DNN is a high-dimensional model of correlations that may capture not only unmodeled dynamics overlooked by physics model but also spurious corrections in the training samples (overfitting). As a result, while DNN has statistically significant performance improvement, but it may also generate outputs inconsistent with the laws of physics that could lead to catastrophic failures. We propose a solution that combines the best of the physical model and DNN model.

Traditionally,  we project a higher dimensional space to a lower dimension space for computational efficiency at the cost of performance. Physics model-based neuron editing turns this conventional projection approach upside down. We project a lower dimensional physics model onto a higher dimensional DNN model by 1) augmenting the input feature vector with physic variables and 2) removing relations permitted by DNN but contradicting the laws physics. The physics model edited DNN  is constrained by physical laws. We gain control-performance at the cost of computing.  To ensure stability of control in spite of DNN software failures, we use Simplex architecture to ensure that the states under the control of the edited DNN must remain within the stability envelope of the physical model based control.

Positions Available:  One Postdoc and part-time jobs in the lab: software development and/or hardware devices for research.

Other Current Projects

• Virtual Sully  
•  Medical Best Practice Guidance (GPS) Systems

Biography: Lui Sha graduated with Phd. from CMU in 1985. He worked at the Software Engineering Institute from 1986 to 1998. He joined UIUC in 1998. Currently, he is Donald B. Gillies Chair Professor of Computer Science, the University of Illinois at Urbana-Champaign. He was named Tau Beta Pi Daniel C. Drucker Eminent Faculty in 2017. He is a Fellow of ACM and IEEE and a recipient of IEEE’s Simon Ramo medal, which honors exceptional achievement in systems engineering and systems science.  IEEE Medals are the highest distinctions that the IEEE presents.

Sha has led the creation of a comprehensive system engineering approach to design and build complex real-time systems, advancing the field from one using hand-crafted, trial-by-error processes into one that is a scientific engineering discipline. The approach, called Generalized Rate Monotonic Scheduling (GRMS) theory, developed with John Lehoczky and Raj Rajkumar, provides predictability, efficiency, and flexibility for scheduling complex concurrent real-time tasks. GRMS has become the best practice of the real-time computing industry, and is regularly taught in real time computing classes. Sha’s IEEE Fellow citation states, “for technical leadership and research contributions which enabled the transformation of real-time computing practice from an ad hoc process to an engineering process based on analytic methods.”

From 2015-2017, Sha was appointed by Administrator of NASA, Charles Bolden, to the Aeronautics Committee of the NASA Advisory Council. He was selected for this council based largely his contribution to real-time computing and to complexity control architecture known as. the Simplex architecture, which allows the safe use of difficult or unverifiable complex control software such as DNN. Its principles have been successfully used to improve the stability advanced avionics systems. Sha and his team also invented the Physically Asynchronous Logically Synchronous (PALS) architecture. Steven P. Miller of Rockwell Collins Inc. demonstrated that, using PALS architecture, the model checking time of a dual redundant flight control system dropped from over 35 hours to less than 30 seconds.  Sha and Miller’s team received the 2009 David Lubkowski Award for the Advancement of Digital Avionics from American Institute of Aeronautics and Astronautics.

 Examples Research Impacts