A rapidly reprogrammable neural network architecture with the possibility for a large synapse matrix is presented. The concept is based on the use of bacteriorhodopsin as a molecular computational element with electrooptical characteristics that are associated with a series of intermediates that are photochemically initiated. One of these states has been stabilized by several orders of magnitude with specific environmental conditions, and this allows the concentration of intermediates to be readily affected without the need for continuous holding illuminations. Thus, the photoelectrical characteristics at each synapse can readily be modulated, and a scheme has been devised to read the synaptic matrix without erasing the impressed synaptic strengths. Electrical measurements are presented to test specific aspects of the overall neural network implementation, and the results of these measurements are encouraging for the development of such a distinctive neural network device.
© 1991 Optical Society of AmericaPDF Article