Mixed signal CMOS high precision circuits for on chip learning

Abstract

Learning algorithms have become of great interest to be applied not only to neural or hybrid neuro-fuzzy systems, but also as a tool to achieve a fine tuning of analog circuits, whose main drawback is their lack of precision. This paper presents accurate, discrete-time CMOS building blocks to implement learning rules on-chip. Specifically, a voltage mode high precision comparator as well as an absolute value circuit. These blocks, plus multiplexing in time techniques, are used to build a circuit to determine the polarity of the learning increments. An exemplary circuit has been simulated with HSPICE with the parameters of a 1 /spl mu/m CMOS technology. Statistical variations of technological parameters were considered. The results show that all curves from 30 runs of a Monte Carlo analysis behave as expected, and at least 8 bits of resolution are achieved by the proposed techniques.