The potential of high speed opto-electronic neural networks

Optics can play an important role in overcoming the bottlenecks imposed on purely electronic systems by the power and space demands of connections between chips. Optical architectures can achieve a high density of interconnection more simply by making use of the third dimension; electronic structures are essentially two dimensional, albeit with multiple levels. Neural networks, because of their dependence on dense interconnection, have the most to gain from the inclusion of optics and many workers have developed opto-electronic implementations. The authors have built a preprogrammed opto-electronic neural network which performs the exclusive-OR operation on two incoming data streams at 50 Mbit/s. It makes use of a 4×4 array of multi-quantum-well surface light modulators. They are fabricated by planar technology, permitting much larger arrays to be made, and have a potential operating speed in excess of 1 Gbit/s. Using these devices, one can obtain large-scale parallelism and high speed, offering new opportunities in neural network design. A significant impact of high speed operation is likely to be on the time-consuming learning phase. Being able to rapidly process a training set of input data will make new training algorithms attractive which are too slow to be practicable with current technology