Interfacing multispectral sensors to real time processors based on neural network models

Publisher Website
Google Scholar
Abstract
Experiments have been carried out to demonstrate and study the interfacing of silicon sensors to silicon devices which represent first stage elements of an artificial neural network. Sensor outputs (network inputs) are photocurrents associated with infrared, visible, or ultraviolet light. First stage linear coding of input current into the pulse rate of a stereotypical neuronlike spiketrain output has been achieved with a dynamic range of more than 106. For 1 pA inputs, the estimated noise referred to input is 10 fA. Network elements are shown to obey equations of the same form as equations which occur in nonlinear neural network models recently analyzed by Hopfield [Proc. Natl. Acad. Sci. 81, 3088 (1984)] and previously studied by Sejnowski [J. Math. Biology 4, 303 (1977)].