Integrated circuit neural networks using ferroelectric analog memory

Abstract

The application of ferroelectric thin-film capacitors as synapses in integrated circuit implementation of artificial neural networks is described. A continuous-valued synapse where the synaptic efficacy is controlled by the polarization of a ferroelectric capacitor element is presented. The ferroelectric analog synapse utilizes nondestructive readout. A ferroelectric capacitor circuit model has been added to the SPICE circuit simulation program and is complete with respect to current vs. time, polarization, and current vs. voltage responses. The simulated response in characterization circuits is shown to closely model that of lead zirconate titanate (PZT) thin-film capacitors. Simulation and experimental results show the proposed synapse circuit to possess a number of favorable characteristics, including very infrequent refresh and fine programming granularity.<>