Fidelity in weighted, volume interconnects [optical neural networks]

In practice, adaptive networks generally require a large number of processing nodes, the output of each being passed, via a weighted interconnection, to many of the other nodes. Thus a network of 10⁴ nodes can require up to 10⁸ individually weighted interconnections. The situation is further complicated by the requirement that independent updating of the weight values is needed during network optimisation. A quantitative analysis is attempted, with a view to assessing the usefulness of an interconnect scheme using multiple grating volume holograms for adaptive network applications. The main question to be answered concerns the fidelity of the system i.e. the degree to which the desired weighted interconnections can be achieved