Study of the effects of nonidealities in multilayer analog neural networks with circuit level simulation

In this study, an analog multilayer neural network was designed using 2 micron SCMOS technology and simulated via SPICE. The simulation results were compared with original training results which were obtained from a neural network simulator, and the discrepancies were investigated. Among the reasons for the discrepancies are limited precision in synapse weights, synapse function non-idealities, and neuron function non-idealities. Simulations were done to study the effects of various degrees of weight precision. Non-ideal operation of synapses and neurons were studied and this behavior was incorporated into the training software to reduce the error resulting from circuit non-idealities.