Title: 

Energy-efficient Digital Hardware for System Identification of Integrated Circuit Systems

Committee: 

Dr. Mukhopahdyay, Advisor

Dr. Raychowdhury, Chair

Dr. Yalamanchili

Abstract: 

The objective of the proposed research is to propose energy-efficient hardware to perform system identification on complex systems, specifically a nonlinear dynamic system. To do this, we first analyze a frequency-domain system identification method of a simple linear thermal system; multi-input and multi-output (MIMO) system. This simple example demonstrates how much system identification is an important problem in engineering domain. Then, we extend the system to be estimated into a nonlinear system, especially image processing (classification) or temporal sequence mapping (power pattern-workload). The system identification of such nonlinear systems can be successfully done by neural networks (feedforward or recurrent). To design energy-efficient neuromorphic hardware, we analyze the impact of hardware-induced error on the performance (accuracy) of several neural networks in either learning or inference. This algorithmic analysis is expanded to demonstrate energy-efficient digital neuromorphic hardware.