Programmable, modular CNN cell

An experimental monolithic implementation of a programmable cellular neural network (CNN) is reported. It overcomes some of the characteristics and restrictions inherent in CMOS VLSI technologies, and allows an arbitrarily large continuous-time analog CNN to be built up by modularly connecting CNN chips with a modest number of cells. The template values are step-wise programmable, with values chosen for functionality rather than according to conventional binary weighting. All external input, output and control signals are electrical and digital, so the CNN can be directly connected to a controller. The design was carried out in a 1-micron n-well CMOS technology. Each cell occupies 0.4mm², including all support circuitry; only one cell per chip was integrated in order to facilitate circuit testing. Measured CNN transients from a prototype 4×4 CNN, formed by connecting 16 one-cell chips are shown. The principal intended applications are the processing of acoustical signals and algorithm development.