Many beneficial applications of computers cannot now be realized because they demand excessively high performance and/or low power. An example of interest is real-time image processing to restore sight to the visually impaired. This project explores an unconventional and little-understood technology called stochastic computing (SC) which is very well-suited to such applications. Stochastic Computing processes numbers in the form of bit-streams that resemble neural signals and are interpreted as probabilities. It can implement complex arithmetic operations by small logic circuits. However, high accuracy may require long bit-streams that can be difficult to interface with conventional binary logic. The project aims to develop a comprehensive theory of SC leading to practical methods for designing and applying stochastic circuits. It will study the foundations of SC, especially speed, accuracy and hardware-cost trade-offs, using various novel methods. Stochastic Computing will be applied to a broad set of image-processing tasks, ranging from retinal implants for the blind, to on-the-fly feature extraction. Prototype designs will be constructed and evaluated using software simulation and hardware emulation via field-programmable gate arrays.
The project's goal is a full theoretical and practical understanding of stochastic computing in the context of emerging integrated circuit technologies and applications. Its results should interest research engineers and scientists in academe, as well as in the microelectronics, computer, and bioengineering industries. They should also be of direct practical value to designers and manufacturers in such application areas as image-processing chips, implantable medical devices, and video surveillance systems. The project's outputs will be distributed primarily via peer-reviewed journal and conference papers. A key goal is to support the training of graduate students in computer science and engineering, who will participate directly in the research as part of their M.S. and Ph.D. programs at the University of Michigan. A few undergraduates will also be invited to join the project as interns to encourage them to pursue research-related careers. A special effort will be made to involve women and minority students.
