Abstract

Because of their speed and large number of spatial degrees of freedom, optics is a promising technology for neural networks. Many proposed optical neural net architectures use free-space interconnects to accomplish the combinations of synaptic weights that are required. One fundamental obstacle to building some such systems is the fact that two coherent optical beams may not be added or subtracted unless they are co-linear. A consequence of this is that many proposed architectures will not work, since they require the addition or subtraction of coherent beams arriving on a detector from different angles.

© 1991 Optical Society of America