Ferroelectric connections for VLSI: Implementation of neuromorphic computation

Neuromorphic computation provides a paradigm which promises the solution of problems in recognition, optimization, and control that have proved elusive using conventional computing. Synaptic connection circuits suitable for the construction of neuromorphic computing systems in VLSI hardware are described. Ferroelectric lead zirconate titanate (PZT) thin-film capacitors, with their electrically alterable polarization, constitute the basic memory elements, with the polarization value representing the stored analog or binary weight. A ferroelectric capacitor device model which have been added to the SPICE circuit simulation program is described and results are shown which demonstrate the validity of the model. Contrary to previously published models and simulators suitable in a limited domain, proper response to any stimulus is manifested. Consequently, the model is appropriate for the simulation of ferroelectric characterization circuits, digital memories, and analog circuits and memories of the type developed herein. Experimental and simulation results are presented which demonstrate that the synapse circuits possess a number of favorable characteristics, including high density, very infrequent refresh, and fine programming granularity. The implementation of simple network architectures and problem solutions is discussed. An introduction to neuromorphic computation, a review of the state of the art in neural computer hardware, as well as an overview of the important electrical aspects of ferroelectric PZT thin-film capacitors and memory circuits are also presented.