Nonstandard A/D conversion based on symmetric neural networks

A technique for nonstandard A/D (analog-to-digital) conversion of signals subject to a fidelity criterion is described that is based on symmetric neural networks. Massively parallel analog arrays offer a natural means for efficient implementation of the proposed technique. Applications include PCM coding, oversampled A/D conversion, and digital image halftoning. A novel kind of parallel analog network (differential neural network) is introduced and shown to be appropriate for nonstandard quantization. These networks contain a nonmonotonic nonlinearity in lieu of the sigmoid function.<>