Analog neural network development system with fast on line training capabilities

Abstract

This paper introduces a system to train a specific neural network device. The system allows the use of standard learning algorithms (e.g. backpropagation, Madaline III, etc.) and generates the weights for a 128 analog neural network device called ETANN (Electrically Trainable Analog Neural Network). An ultra fast weight setting algorithm stores the internal coefficients into a real analog (not computer simulated) neural network. The trainer and programmer (NETMAP) system is able to cycle through steps of learning, weight setting and parallel processing runs, performing the so called chip-in-the-loop procedure. The NETMAP system integrates hardware and software to interface and take full advantage of ETANN parallel processing capabilities. Due to weight setting is a complex and computationally intensive process, a solution, based on an off-line mapping between weights and pulses of charge, is proposed to speed up the loading time. >