Mixing analog and digital techniques for silicon neural networks

Some of the problems which arise during the design of dedicated implementations of artificial neural networks are described, with emphasis on the selection tradeoffs between analog and digital circuits. It is shown that mixing both techniques may result in an overall performance improvement. A cell based on this principle is described. The analysis of several implementation problems of silicon ANNs (artificial neural networks) shows that precision is not essential for adaptive networks and that pulse-stream circuits represent a good compromise between performance, size and speed. Pulse-stream methodologies are chosen because they allow a noise-insensitive, hence efficient, transmission of information. Although this has not been the selection criterion, it is worth noting that the use of pulse streams has a counterpart in natural neural networks, which use these techniques heavily to transmit analog information in a noisy environment and to reduce the total power dissipation