Soft competitive learning using stochastic arithmetic
Brown, Bradley D.
Artificial Neural Networks (ANNs) are massively parallel systems which can benefit from technology which allows implementation of an unusually large number of simple computational elements on a single integrated circuit. Unlike traditional computation devices such as microprocessors, neural networks are also characterized by a tolerance for much less accurate computation. Stochastic arithmetic may be performed by computational elements which are both very small and compatible with modern VLSI design and manufacturing technology. This thesis presents a number of stochastic computational elements, several of which are introduced for the first time in this thesis, and an analysis of their operation. The applicability of stochastic arithmetic to neural networks is demonstrated through the successful implementation of a sample problem, optical character recognition, using stochastic computation. While the accuracy, power and speed characteristics of stochastic computation may not compare favorably with more conventional binary radix based computation, the low circuit area requirements make them attractive for VLSI implementation of ANNs. Results are presented for an example ANN application. Optical character recognition is performed on the characters in the E-13B MICR (Magnetic Ink Character Recognition) font. (Abstract shortened by UMI.)