This project proposes a new paradigm to enhance computing efficiency --- Finite State Machine (FSM)-centric approximate computing. Approximate computing has shown promise for both reducing energy consumption and improving performance across different applications, especially those in image processing, machine learning and data analytics. To date, approximate computing has been inapplicable to FSM modeling of computations, which has important applications in domains that include biological science, cyber security, data compression, software engineering and hardware design. Growing data volumes and limitations on computer processing power constrain FSM?s efficiency. The establishment of FSM-centric approximate computing will open the door to a new dimension of efficiency optimization for software applications.
This research will take advantage of the synergy between FSM computations and approximate computing --- the inherent error tolerance capability within FSM computations --- to develop a computing platform for exploring approximate FSM computations. The key idea is a quantitative analysis of FSM reliability that captures how errors generated by underlying approximate hardware propagate through FSM transitions. Additionally, this research will also design and implement two complementary approximation schemes --- one relies on the "inexactness" of approximate hardware; the other provides pure software approximation and runs on conventional exact hardware. Together these approximation strategies will demonstrate the potential of FSM-centric approximate computing in improving the efficiency of FSM applications.