A programmable intelligent optical sensor realization

A hybrid VLSI realization of a programmable, multi-layer optically coupled neural network is described in this paper. The network itself is a fully connected, multi-layer feedforward network. It has an optical array constructed by 5/spl times/5 photosensitive elements (photoreceptor) as the input devices. The artificial neurons have been implemented by 2-stage comparators composed with differential amplifier and output inverter to model the signum (hard limiting) nonlinear function. Scaled currents generated by the cascode current mirrors are selected by the data stored on the on-chip dynamic shift register (DSR) and performed as the synaptic weights. There are 25 input nodes, 4 hidden neurons and 3 output neurons for this network. The physical realization which contains approximately 9281 devices has been designed and ready to be fabricated by 1.2 /spl mu/, double-metal, n-well full custom CMOS technology.