Processing noisy analog signals from microsensor arrays and VLSI imagers using stochastic binary computations

This paper describes the use of Bernoulli-distributed binary random variables to implement signal processing from integrated arrays of parallel microsensors and VLSI imagers. Parallel digital signal processing of these signals is achieved in hardware by a combination of CMOS threshold detection of noisy raw signals, and the use of stochastic arithmetic (see Gaines, B.R., "Advances in Information Systems Science", Plenum Press, vol.2, chap.2, p.37-172, 1969; Brown, B.D. and Card, H.C., IEEE Trans. Computers, vol.50, p.891-905, 2001). Hardware efficiencies include reduced power dissipation and improved error tolerance, as compared to binary radix implementations, together with the ability to trade precision against computation time using fixed hardware, reminiscent of biological computations.