A synthetic neural integrated circuit

Abstract

Integrated circuits are approaching biological complexity in device count. Biological systems are fault tolerant, adaptive, and trainable, and the possibility exists for similar characteristics in ICs. The authors report a limited-interconnect, highly layered synthetic neural network that implements these ideals. These networks are specifically designed to scale to tens of thousands of processing elements on current production size dies. A compact analog cell, a training algorithm, and a limited-interconnect architecture which has demonstrated neuromorphic behavior are described